ATI Rage 128 Roundup – A Complete Review Of The Series

  By ,

Let me invite you to review a series of graphics cards that is often overlooked. There are too many variants of similarly named cards, (many of which are painfully slow), too many stories of terrible drivers and image quality issues, and an absolutely absurd number of model names that nobody can make sense of.

Most retro gamers avoid ATI's Rage 128 series like the plague and opt for nVidia or 3dfx cards instead.

I have made it my mission to shed light on this series of cards once and for all, and clear up all the confusion surrounding them. In the following article, which ended up being much more extensive than I planned, I will discuss general aspects such as technical differences, 3D features, image quality and historical considerations.

At the end of the article (as a hint for the impatient among us), there are also two very interesting things to be found:

  • A table containing virtually all Rage 128/Pro model numbers including key technical data. For each model number, there is also a link to in the table, where all relevant data for the card can be found in greater detail.
  • A huge number of benchmarks performed on three different systems, as well as a complete performance ranking for all model numbers on each system.

But before we get started, a quick warning: this is a very long article, so if you are interested in these graphics cards beyond the benchmark values, you should really take your time. I'm sure some of the information will surprise you, and there is a lot to discover.

First Generation: Rage 128 GL

The Rage 128 GL was ATI's entry for the 2nd generation of 3D chips. It was meant to compete with nVidia's TNT, the Matrox G200, S3 Savage 3D and 3dfx' first all-in-one 3D chip - the Voodoo Banshee. Out of all these, it was announced the latest (August of 1998) and, because of stability problems, it was not before January of 1999, when the first cards finally hit the market.

Like S3 with its Savage 3D, ATI decided to use the then-modern 250nm manufacturing process for their Rage 128 chips, while 3dfx, Matrox and nVidia still went with the proven 350nm manufacturing process. And like other companies before adopting this new technology, ATI was facing stability issues with their product. They had to revise the chip to get it review-stable, before pre-production samples could be shipped to reviewers in early December of 1998 - more than three month after the chip was originally introduced. And still, the initially proclaimed 100 MHz clock for the core had to be reduced to 90 MHz on the first review cards, to ensure stable operation.

Because of the finer manufacturing process, the chip should generate less heat and, furthermore, the original datasheet says it can operate safely up to a temperature of 95 degrees Celsius (125 maximum "in storage"). This might have been the reason why ATI decided to ship these review cards without a heatsink. This was criticised by some reviewers, who reported temperatures of more than 90 degrees during their tests.

Further issues in early reviews included stability problems with Super Socket 7 systems, image quality problems in older DX5 based titles and several driver-related problems - mostly in OpenGL based games, since, at the time of these reviews, there was only a Beta OpenGL ICD available.

But in general, the chip got mostly positive reviews and was praised for its unparalleled performance in 32-bit color depth, which was often not even 10 % lower than its 16-bit numbers, leaving nVidia's TNT chanceless in 32-bit benchmarks. It also fared well with lower-spec CPUs, because of its powerful triangle setup engine - something the TNT was missing alltogether.

The A22 Silicon

    But by the time of the card's first reviews in December 1998, ATI was still troubled by unstable chips and thus low production yields. For the release to market in early January of 1999, ATI decided to apply some further changes to the production cards with the new A22 chip-revision. In an internal "Product Advisory" document from January of 1999, several measures are mentioned:

      • VCore increase from 2.5V to 2.8V
      • usage of a passive heatsink
      • maximum frequencies (mem/core) of 125/80 for the VR and 100/90 for the GL (though early Rage Fury cards with 103/103 exist)
      • usage of an oscillator instead of a crystal for the generation of the reference frequency (the crystal was identified as a source of noise on internal clocks)
      • noise filtering on core power supply (the first batch of cards all have MLCCs on the backside of the chip for this reason)
      • change of a specific resistor value (RSET)

      Technically, this makes the Rage Fury/Magnum more a 1999 card. But almost everywhere in the net, August of 1998 is mentioned as the release date of the Rage 128 GL chip. While I cannot rule out the possibility that a few cards were already shipped in December of 1998, the earliest PCB dates I could find were from the first weeks of 1999. Furthermore, I have read in several old magazines that retail sales, at least in Germany, did not start until March 1999.

      However, as far as I am aware, the chips found on the review cards in late 1998 were identical to those shipped to customers, so I am listing the release date of the chip as 1998-11. Furthermore, in the "Rage 128 VR and 128 GL Graphics Controller Specifications" document from November 1998, the Rage 128 GL chip is listed as 215R4GASA22, which already corresponds to the revised A22 variant of the chip. The Rage 128 VR chip on the other hand is still being listed as 215R3BASA21, which is the A21 silicon.

      Rage 128 VR: ATI’s Budget Option

      The 128 VR chip was introduced to cover the budget market and was identical to the Rage 128 GL in terms of features, only its memory interface was halved to 64 bits. Although VR cards had a higher memory clock (up to 125 MHz depending on the model), the memory bandwidth was much lower than that of GL cards, due to its 64-bit memory interface. Additionally, the core clock was reduced to 80 MHz.

      The VR chip exists in two packaging variants: a compact 256-pin BGA package for integration on motherboards and a 272-pin BGA package for integration on graphics cards. The 272-pin chip has the same physical size as the 328-pin GL chip, only the latter has several more pins towards its centre.

      However, the whole pinout is different. Why am I pointing this out? Well - the first "VR" cards (109-52xxx-xx) really had a VR chip installed, later "VR" cards (109-66xxx-xx) feature a full Rage 128 GL chip, with only half its memory interface connected. The difference is quite obvious when we compare the area around the chip on the backside of four different VR cards:

      109-52000-00

      109-52100-01

      109-66500-00

      109-68100-01

      It is easily visible, that both 52xxx cards look similar and so do the 6xxxx cards compared to each other. But the PCB layout around the chip differs quite a bit between 109-52xxx-xx and 109-6xxxx-xx cards.

      Furthermore, the latter report the same ASIC ID as cards with the GL chip (5246 Rage 128 GL AGP 2x), whereas the 109-52xxx-xx cards report their ASIC ID as 524c Rage 128 RL/VR AGP 2x. Keep this information in mind for later on, when we take on the Rage 128 Pro series of cards.

      Pricing

      The 16 MB variant of the GL-based cards (Xpert128) was a bit more expensive than the Riva TNT and Voodoo Banshee in Germany in April 1999. While you had to pay 225 DM (~110 USD) for the ATI card, Riva TNT cards from various manufacturers were available for between 195 DM and 209 DM (e.g. Hercules, Creative, Yakumo, STB). A Voodoo Banshee was even cheaper at between 165 DM (Yakumo) and 189 DM (Asus/Creative). Unfortunately, I couldn't find any retail prices for the 32 MB Rage Fury for this period.

      Its rather high price and delayed availability shure give a hint why there are very few retail variants of these cards (easily identified by the S-Video and composite-out ports on a little extra-PCB) available on the used market and why the card played practically no role on the retail market back in the day.

      Quite in contrast to the OEM market, where Rage 128 GL and VR models were sold in large numbers. This explains why so many cards with these chips are still readily available today. Not all of them guarantee a worry-free retro gaming experience, though. This is why we will try to examine "all" variants of these cards in terms of performance towards the end of this article, so you can easily identify if the Rage 128 card(s) in your collection is/are suitable for retro gaming or a waste of storage space.

      Technical Specifications Compared To Competing Cards

      Model Rage Fury Rage Magnum Xpert 128 Xpert 99
      Manufacturer ATI ATI ATI ATI
      3D Chip Rage 128 GL Rage 128 GL Rage 128 GL Rage 128 VR
      Release Date (Chip) 1998-11 1998-11 1998-11 1998-11
      Manufaturing Process 250 nm 250 nm 250 nm 250 nm
      Core Clock 103 MHz 90 MHz 90 MHz 80 MHz
      Mem Clock 103 MHz 90 MHz 90 MHz 125 MHz
      Fill Rate (MPix/s) 206 MPix/s 180 MPix/s 180 MPix/s 160 MPix/s
      Fill Rate (MTex/s) 206 MTex/s 180 MTex/s 180 MTex/s 160 MTex/s
      Memory 32 MB SDR 32 MB SGR 16 MB SDR 8 MB SDR
      Mem Bus Width 128 Bit 128 Bit 128 Bit 64 Bit
      Mem Transfer Rate 1648 MB/s 1440 MB/s 1440 MB/s 1000 MB/s

       

      Model Asus V3400TNT Asus V3200 Sparkle SP341E Matrox  G200
      Manufacturer nVidia 3dfx S3 Matrox
      3D Chip Riva TNT Banshee Savage 3D G200
      Release Date (Chip) 1998-09 1998-10 1998-08 1998-07
      Manufaturing Process 350 nm 350 nm 250 nm 350 nm
      Core Clock 90 MHz 100 MHz 100 MHz 85 MHz
      Mem Clock 110 MHz 110 MHz 100 MHz 113 MHz
      Fill Rate (MPix/s) 180 MPix/s 100 MPix/s 100 MPix/s 85 MPix/s
      Fill Rate (MTex/s) 180 MTex/s 100 MTex/s 100 MTex/s 85 MTex/s
      Memory 16 MB SGR 16 MB SGR 8 MB SDR 8 MB SGR
      Mem Bus Width 128 Bit 128 Bit 64 Bit 64 Bit
      Mem Transfer Rate 1760 MB/s 1760 MB/s 800 MB/s 904 MB/s

      Second Generation: Rage 128 Pro

      In early October of 1999, around ten months after the Rage 128 was readily available to customers, ATI launched its successor - the Rage 128 Pro. Once more, ATI was the latest to the party, as S3, 3dfx, nVidia and Matrox all had their new products already on the shelves. nVidia even released their second new chip for 1999 (the Geforce 256) only a few days later. The same goes for S3's Savage 2000, which got released in November of that year.

      The MSRP of the Rage Fury Pro was at 149 USD; the earliest listings I could find for Germany were around 390DM in December 1999, which translates roughly to a little less than 200 USD for that time. At the same time a "cheap" TNT2 Ultra could already be had for 399DM (most models were still priced towards 450 to 500 DM, but prices dropped with the launch of the GeForce 256).

      The Matrox G400 DualHead with 32 megs was comparably priced at 400 DM. Its SingleHead variant, as well as 3dfx' Voodoo 3 3000 were only around 330 DM by the end of 1999. This means the Fury Pro was facing similar problems as its predecessor: late release, comparably fast or even faster cards could be had cheaper, next generation cards of competitors were (very) soon to be released.

      What’s New?

        Now let's turn our attention to the improvements under the hood: what exactly did ATI do to squeeze out more out of the Rage 128 architecture?

          • The AGP transfer speed was raised from 2X to 4X (for these class of cards this is more a "cosmetic" change, the major advantage is compatibility with more modern systems)
          • Support for DX6 texture compression (allegedly; we'll cover this later on)
          • Drastically improved triangle setup engine that increases the performance from 4M triangles/sec to 8M triangles/sec (this helps in polygon-heavy games / scenes)
          • ~20% increase in clock speed (depending on model)

          Another improvement the Rage 128 Pro is supposed to have (according to several sources around the net), is the support for anisotropic filtering. But this is definitely false information, which we'll come to later when we also take a look at the support for texture compression. Anisotropic filtering was mentioned as an upcoming feature in an early "leak" for example, but did not make into "ATI silicon" until their first Radeon card.

          Variants

          In terms of chip- and memory clocks, the second Rage 128 Generation is (at least a little) more straightforward. By the time the chip launched, there were excactly two variants of cards available (not taking secondary features like TV-In/-Out into account), both of which could be had with either 16 or 32 MB of RAM:

          Card Rage Fury Pro Rage Fury Pro
          Chip Rage 128 Pro Rage 128 Pro
          Chip Clock 118 MHz 118 MHz
          Mem Clock 140 MHz 140 MHz
          Mem Size 16/32 MB 16/32 MB
          Mem Type SDR SGR
          Mem Bits 128 128

          It is worth mentioning, that there are a lot of Fury Pro cards out there, which clock at 120/120. These are variants without video-in functionality, which could still feature various combinations of composite and s-video outputs). The lower clocked cards all come from OEM machines, while cards with both of these outputs plus composite-in were found in retail boxes and clock at 118/140 MHz.

          Around three months later, the XPERT 2000 Pro was launched to cover the lower end of the market. It featured the Rage 128 Pro VR chip, 32 MB of SDR 64-bit memory and operated at clocks of 120/120 MHz core/mem. As far as I am aware, the XPERT 2000 Pro was also meant for retail sales, as competition to cards featuring nVidia's TNT2 M64 chip, but I have never seen a retail box for it. What I can clearly remember is reading XPERT all over advertisings for OEM machines back in the day. This is were most of these cards went, continuing ATI's success in the OEM market.

          But there is only one card, which uses the Pro VR chip - the L-shaped 109-65700-xx. Remember when I said "keep this information in mind" in the Rage 128 VR subsection before? The first-gen VR has a different ASIC ID, which helps identifiying cards with this chip. The Rage 128 Pro VR on the other hand reports itself as 5046 Rage 128 PF/Pro AGP 4x (TMDS), which is the same as a regular 128 Pro chip.

          My first conclusion was, that there is no such thing as a 128 Pro VR chip. Unfortunately, first-hand information (official datasheets/documents directly from ATI) on the Rage 128 Pro generation of chips is very sparse, so I cannot provide a definitive answer as to whether a separate Rage 128 Pro VR chip exists or if these cards also have a 'full Pro' chip installed.

          But looking on the backside of the L-shaped card, the area surrounding the chip differs from all other Pro models, while the remaining models look pretty similar or even identical:

          109-60600-10

          109-63200-01 (16MB)

          109-63200-01 (32MB)

          109-65700-20

          The first 3 cards look very similar or even identical. The fourth picture shows the backside of the L-shaped 109-65700-20 card, which looks vastly different around the chip.

          Although I couldn't find a pinout for the Pro chip or first-hand information on whether a separate VR chip exists, I am assuming that the latter is the case, as it normally makes sense to have a cost-reduced chip for OEM sales.

          Technical Specifications Compared To Competing Cards

          Model Rage Fury Pro Retail Rage Fury Pro OEM Xpert 2000 Pro
          Manufacturer ATI ATI ATI
          3D Chip Rage 128 Pro Rage 128 Pro Rage 128 Pro VR
          Release Date (Chip) 1999-10 1999-10 2000-02
          Manufaturing Process 250 nm 250 nm 250 nm
          Core Clock 118 MHz 120 MHz 120 MHz
          Mem Clock 140 MHz 120 MHz 120 MHz
          Fill Rate (MPix/s) 236 MPix/s 240 MPix/s 240 MPix/s
          Fill Rate (MTex/s) 236 MTex/s 240 MTex/s 240 MTex/s
          Memory 32 MB SDR/SGR 32 MB SDR/SGR 32 MB SDR
          Mem Bus Width 128 Bit 128 Bit 64 Bit
          Mem Transfer Rate 2240 MB/s 1920 MB/s 960 MB/s
          Model Diamond Viper V770 Millennium G400 DH Voodoo 3 3000 PowerColor CM64A
          Manufacturer nVidia Matrox 3dfx nvidia
          3D Chip Riva TNT2 G400 Voodoo 3 TNT2 M64
          Release Date (Chip) 1999-04 1999-05 1999-04 1999-10
          Manufaturing Process 250 nm 250 nm 250 nm 250 nm
          Core Clock 125 MHz 125 MHz 166 MHz 120 MHz
          Mem Clock 150 MHz 166 MHz 166 MHz 125 MHz
          Fill Rate (MPix/s) 250 Mpix/s 250 Mpix/s 166 MPix/s 240 Mpix/s
          Fill Rate (MTex/s) 250 MTex/s 250 MTex/s 333 MTex/s 240 MTex/s
          Memory 32MB SDR 32 MB SGR 16 MB SDR 32 MB SDR
          Mem Bus Width 128 Bit 128 Bit 128 Bit 64 Bit
          Mem Transfer Rate 2400 MB/s 2656 MB/s 2656 MB/s 1000 MB/s

           

          What About Pro2 / Ultra?

          Later in the Rage 128 Pro's lifespan, newer chip revisions got released, some of which are also listed as "Rage 128 Pro2". These later revisions are mostly found on OEM "Rage 128 Ultra" cards, which can be anything between the slowest and the fastest Rage 128 Pro cards you can get (with the latter being the exception).

          There are several reports of people having issues with these Pro2 / Ultra cards. In some cases, the latest beta driver seems to have worked when everything else failed. I can not recommend this driver though, as especially its OpenGL performance is among the worst of all Rage 128 Pro drivers ever released and the D3D performance is also rather poor. If you still want to give it a try, you can find it on soggi's useful site: https://soggi.org/drivers/ati.htm#Rage128

          During my tests, I tried out all 109-7xxxx-xx models featuring the 'Pro2' chip, and can confirm that they work flawlessly with any driver supporting the Rage 128 Pro. Even the oldest ("Pro") driver I found (4.11.6263 from February 2000) worked without issues on my Super Socket 7 (Gigabyte GA-5AX) and Socket A (EPoX 8KTA3) machines. However (!), I experienced serious issues on my Slot1 machine (Asus P3B-F) with every single driver I tried.

          These later "Ultra" cards definitely have compatibility issues with certain boards, but work flawlessly with others. My advice is: if your card doesn't install right away, or if your system behaves strangely during or after installation of the driver (e.g. black screen after boot, freezing, screen distortions, bluescreens), switch to the next platform or board immediately. You won't be able to get these issues sorted out.

          In addition to the Ultra cards with ATI product numbers (built by ATI), there are several 3rd party cards featuring  this chip (HIS, Sapphire, Manli,...), all of which are low-cost and low-performance cards. Furthermore, the driver issues will be even worse, so definitely stay away from those.

          The Pro2 Difference

          The "Pro2" Chip is one of the biggest (if not the biggest) question-marks in the Rage 128 story. I couldn't find any official documents or news about this chip, which added to my initial belief that the Pro2 chip is simply a later revision of the Pro, which can handle higher clocks due to optimized production. But then again...

          Compatibility Issues With Certain Motherboards

            Like I already mentioned in the previous part, I couldn't get any Pro2-based cards to work with the motherboard in my Slot 1 benchmarking system, which is the Asus P3B-F. I experienced similar issues to those reported by many people online as driver-related problems with these cards. However, these issues are not actually driver-related, but rather come from incompatibilities with certain motherboards. Unfortunately, I cannot provide you with guidelines on which chipset, platform or motherboard manufacturer might be problematic, as my testing did not reveal a pattern that would help to identify the root cause of these problems.

            Initially, I thought these cards might cause issues in combination with the Intel i440BX chipset that the Asus P3B-F is based on, since my other two benchmarking systems (AMD K6 / K7 platforms) worked without issue. I therefore grabbed my Gigabyte GA-BX2000 for testing, and interestingly, I had no problems installing the 109-73100-02 and running some benchmarks.

            My next best bet was, that the problem might only occur on Intel BX boards from Asus, so I tried out my Asus CUBX-E next. But this constellation also proved to be completely trouble-free.

            I ended up with five out of six motherboards working with any Pro2-based card and and any driver supporting the Rage 128 Pro in my small compatibility test, only the P3B-F wouldn't work. Contrary to my hopes and expectations, I wasn't able to identify a pattern which could help avoiding incompatible motherboards from the get-go, but at least I can share an overview of the boards I tested:

              Board Platform Chipset Result
              Gigabyte GA-5AX Super Socket 7 ALi Aladdin V OK
              Asus P3B-F Slot 1 Intel i440BX not working
              Gigabyte GA-BX2000 Slot 1 Intel i440BX OK
              Asus CUBX-E Socket 370 Intel i440BX OK
              EpoX EP-8KTA3 Socket A VIA KT133a OK
              Asus P4S533 Socket 478 SiS 645DX OK

               

              Different ASIC ID

              The ASIC ID identifies the chip that is installed on a card. Normally, it is possible for a revised chip to still have the same ASIC ID, but usually you won't find different ASIC IDs for the same chip. Even in ATI's infamous Rage Pro AGP 2x / Rage Pro Turbo stunt, the ASIC IDs were identical between the chips, despite them having different labels and part number codes, and the "Pro Turbo" being marketed as a new chip.

              However, the ASIC IDs for the Rage 128 Pro and Pro2 are actually different and are as follows:

              • ASIC ID Rage 128 Pro: 5046 Rage 128 PF/Pro AGP 4x (TMDS)
              • ASIC ID Rage 128 Pro2: 5246 Rage 128 GL AGP 2x

              Different Part Number Code

              The newer chip also got a different part number code, which we'll have a look upon in a second. At this point, let me bring up the 128 GL chip again. Its "A22" revision finally got the chip stable and was the revision that was shipped to customers. But what is the difference between A21 and A22? And how does this information help identify, what could have changed with the Pro2 chip? A closer look at the chip codes for GL, Pro and Pro2 might prove helpful:

                  • 215R4GASA22 = Rage 128 GL
                  • 215R4GAUC21 = Rage 128 Pro
                  • 215R4GAQD22 = Rage 128 Pro2

                  As you can see, only the last four positions of the code differ. Let's explore what this all means by decoding ATI's part number scheme.

                  The first three positions (215) indicate, that this is a desktop chip. 216 would be a mobile chip, 213 is used for Rage Theater chips. R4G indicates the model and was initially used for the 128 GL. This part was recycled for the 128 Pro(/2) - another hint on how similar they really are. The next character (A) stands for the substrate revision, which did not change throughout the lifespan of the Rage 128 series and remains the same for all Rage 128 series chips.

                  Now that we covered the similarities, let's move on to the differences. The fourth-last position reveals where the chip was manufactured. Since my information is drawn from a Rage 128 GL datasheet, the list of factories appears to be incomplete ("Q" is missing in the list). Nevertheless, here are the possible letters for this position:

                    • C - SGS
                    • J - UICC
                    • S - TSMC
                    • T - TI Japan
                    • U - UMC
                    • Q - ??

                      The next position is probably the most interesting, as it indicates the feature revision (generation) of the chip. The Rage 128 GL is the first generation (A), "B" is probably some engineering stage (wild guess) and "C" is the Rage 128 Pro. Since the "Pro2" chip has the letter "D" on this position, there might really be a difference in its feature set, a performance optimisation in hardware or a hardware fix. The remaining, unanswered question is what it actually is. During my tests and research, I couldn't find any strinking evidence of performance improvements or new features.

                      The second-last position indicates the "base layers revision" and this, much like the substrate revision, did not change on any Rage 128 based desktop chip throughout its lifespan. On the last position we find the "metal layers revision", which is the second interesting aspect in the evolution from the Pro to the Pro2 chip. First of all, I would like to point out that, since the GL and the Pro are different chips, they also have separate increments on their respective metal layer revisions. Although the GL chip was already manufactured in the improved A22 revision, the Pro started with revision C21, with the trailing "1" indicating the first "metal layers revision" of this new chip.

                      The Pro2 already runs on "metal layers revision" two (revision D22), which enables it to clock higher and, at the same time, run cooler. This is the one key difference one can really grab; quite in contrary to the newer feature revision. To give you at least some approximate data on the temperature aspect, I measured the temperature on the backside of the chip of a Rage Fury Pro (109-63100-11, 118/140 MHz) and a Rage 128 Ultra (109-73100-02, 133/133 MHz) after rendering ten minutes of the Unreal intro (castle flyby loop). And these are the results using a cheap infrared laser thermometer:

                      • Rage Fury Pro: 62.0 °C
                      • Rage 128 Ultra: 58.5 °C

                       

                      Even though it is clocked higher and the size of the heatsink is the same, the Rage 128 Ultra stays 3.5°C cooler. This may not seem like much at first, but it still dissipates around 6% less heat, even though the core clock is around 13% higher.

                        In A Nutshell

                        • Pro2 chips run cooler
                        • Pro2 chips can be clocked higher
                        • Pro2 chips have compatibility issues with certain chipsets (i.e. Asus P3B-F), which look & feel like driver compatibility issues

                        Clock-By-Clock And AGP Bandwidth Comparison

                        To conclude our examination of the hardware side of the Rage 128 series, let's take a quick look at how the 2(.5) generations of chips differ in terms of AGP bandwidth and clock-for-clock performance.

                        A member of the VoodooAlert forum inspired me to draw a comparison between the AGP bandwidth of the Pro and Pro2 chips. As a result of some 3DMark2001SE results he got in his own tests, he got the impression that the Pro2 chip may have performance issues which the "regular" Pro does not have, when the AGP transfer rate is reduced to 2x.

                        Testing The AGP Bandwidth

                        By the time I did the following comparison, I hadn't done more testing with the Pro2 on Slot1 / Intel BX boards - this is why I only have numbers for the VIA KT133A (as a period-correct chipset) and the SiS 645DX (as a more modern platform). The Intel BX would have been an interesting chipset, as it is AGP2x-only. But I can spoil you with the following: The Pro2 performs just as good on the BX with AGP set to 2x as the Pro does. The bandwidth results would have told the same story.

                        To make things more complete, I also included the Rage 128 GL (which only supports AGP2x) and did the test with two different driver versions. The AGP bandwith test was performed using FinalReality version 1.01.

                        AGP Bandwidth VIA KT133A
                        4.12.6292 Chip 4.13.7192
                        173,23 MB/s 128 GL 204,11 MB/s
                        190,20 MB/s 128 Pro AGP2x 205,41 MB/s
                        189,10 MB/s 128 Pro AGP4x 208,17 MB/s
                        187,22 MB/s 128 Pro2 AGP2x 211,99 MB/s
                        189,82 MB/s 128 Pro2 AGP4x 212,85 MB/s
                        AGP Bandwidth SiS 645DX
                        4.12.6292 Chip 4.13.7192
                        212,59 MB/s 128 GL 206,91 MB/s
                        217,21 MB/s 128 Pro AGP2x 218,56 MB/s
                        241,52 MB/s 128 Pro AGP4x 239,53 MB/s
                        180,63 MB/s 128 Pro2 AGP2x 217,39 MB/s
                        185,82 MB/s 128 Pro2 AGP4x 240,10 MB/s

                        The differences aren't dramatic and certainly do not affect the performance of any of the chips on these platforms. For example, the AGP bandwidth achieved on the Intel i440BX is only between 120 and 130 MB/s for the GL and Pro, yet the cards perform almost identically in GPU-heavy scenarios in a BX-based system with a 850 MHz Pentium III than in a KT133A-based system with an Athlon 1400C, despite the latter being roughly 70% faster. Just be aware of that; we'll come to very detailed performance results at the end of this article.

                        Futhermore, while the AGP bandwidth improved for the most part for each chip with the newer driver, the performance of the latest reference driver (4.13.7192) in games, especially using Direct3D, is significantly worse that it is with the older one. This means the improvements are more cosmetic, as the latest driver is not really a viable option in regards to performance in games.

                        Comparing The Rage 128 GL / Pro / Pro2 Clock-By-Clock

                        The next aspect might be a bit more interesting. The Rage 128 Pro is certainly faster than its predecessor and runs at clocks that a regular, unmodded Rage 128 GL cannot achieve. But what if both generations were clocked the same? And does the Pro2 differ in performance from the regular Pro?

                        To find out exactly that, I chose a card with the same specifications (32MB SD-RAM, 128-bit memory interface) for each generation and set the clock speeds to 110/110 MHz. For the Rage 128 GL, these are quite hefty clocks, especially for the chip. I had to use a 120mm fan to cool down the card while benchmarking, otherwise I would get minor graphical glitches after a short while. But with the extra cooling applied, the card ran fine during the tests.

                        All benchmarking was done on my Socket A benchmarking system, using driver version 4.12.6292. I chose Quake 2 and MDK2 for the test, as they both consistently deliver highly reproducible results.

                        Quake 2 demo1.dm2, 1024x768x32
                        109-51900-01 (R128GL, 32MB SDR, 128Bit) 48,1
                        109-63200-01 (R128Pro, 32MB SDR, 128Bit) 48,6
                        109-73100-02 (R128P2, 32MB SDR, 128Bit) 48,5
                        MDK2 Integrated Benchmark, 1024x768x32
                        109-51900-01 (R128GL, 32MB SDR, 128Bit) 28,96
                        109-63200-01 (R128Pro, 32MB SDR, 128Bit) 29,21
                        109-73100-02 (R128P2, 32MB SDR, 128Bit) 29,24

                        To be honest, I already expected the results to be very close, as the only striking differences between the chips are the more powerful triangle setup engine and the increased AGP transfer speed that the Pro has over the GL. As discussed before, the AGP bandwidth is of relatively minor influence on the performance. Therefore, I would have expected the difference in MDK2 to be the greatest, as it is a much newer game with more complex geometry than found in the over two years older Quake 2. This should have been a good game to show off the advantage of the twice-as-powerful triangle setup engine.

                        But this didn't happen, the results in MDK2 are extremely close. There is a very small advantage for Pro and Pro2 over the GL in that game, but even less so than in Quake 2. I think it is safe to conclude, that the raised clockspeeds were by far the biggest contribution to the huge leap in performance of the second Rage 128 generation. I also think we can summarize that the Pro and Pro2 perform the same within the margin of error.

                        Myth Busting: Feature Comparisons and Tests

                        Now that we have covered the background information and the hardware side of the two generations of chips, let's explore the feature set of the Rage 128 GL and Pro - I am shure this part contains some suprises for the very most of you.

                        I made a feature report of each API for both chips for every major driver version, using an older version of Everest Ultimate (thanks for the tip goes to vlask from VKA Legacy MKIII!) and created one big table (actually two; seperated in D3D and OpenGL) for a complete overview of supported 3D features throughout the driver versions. Why so many driver versions? Well, bug-fixing isn't the only thing that happens in newer driver versions, as can be seen impressively in the below tables. Especially the OpenGL part of the driver got a lot of additional features over time.

                        Unfortunately, these tables got too big to integrate them in here natively, so I had to insert them as images.

                        Direct3D Feature Table

                        I highlighted Anisotropic Filtering and Texture Compression, two features that the Rage 128 Pro is supposed to support. Both are reported as "not supported" by every driver - we will have a closer look a these in a second. Interestingly, other features such as Edge-Anti-Aliasing and Mipmap LOD Bias Adjustments were removed in later driver versions. These features are only supported by the Rage 128 Pro when using driver version 4.11.6263, as they were removed in subsequent versions and the aforementioned driver was the first to support the new chip.

                        The only additional D3D features that the Rage 128 Pro supports over the Rage 128 GL are Substractive Texture Blending and Texture Mirroring, the rest of the D3D feature set is identical for both Rage 128 generations when using the same driver version.

                        OpenGL Feature Table

                        Again, Anisotropic Filtering and Texture Compression are highlighted, to make them easier to spot. Like in D3D before, each driver version reports these features as "not supported". I won't bore you with going over every single entry, but it is clearly visible across the various versions, that ATI's driver team put a lot of work into the OpenGL driver, especially after the Rage 128 Pro got supported with version 4.11.6263 and up.

                        By the way - this feature table applies to both the Rage 128 Pro as well as the Rage 128 GL. Ultimately, this means that both chips have an identical OpenGL feature set and differ only very slightly in terms of D3D features!

                        Feature Test: (Anisotropic?) Filtering

                        As I mentioned earlier on, the Rage 128 Pro's claimed support for anisotropic filtering is false information that keeps being rewritten and retold over and over, because it got mentioned as an upcoming feature in early leaks. The above tables clearly show that no driver reports this feature as supported and there is no driver that provides an option to turn this feature on or off.

                        To have definite proof, I installed the Max Payne Demo and created a savegame in the subway, where the difference between trilinear and anisotropic filtering should be clearly visible. May Payne is one of the earliest games, that implemented the option to activate anisotropic filtering directly from the game's option menu. I tried every driver version from 4.11.6263 upwards, but none showed any visual differences between trilinear and anisotropic filtering - as expected, I may add.

                        Bilinear filtering on the left, trilinear filtering on the right. The difference in filtering can be identified best by following the red line on the floor. Use the pillars as an orientation. Like expected, the anisotropic filtering option did not have any effect and looked just like the trilinear filter seen active on the right image.

                        But there's more exciting stuff to discover in terms of filtering: Another thing that keeps on being rewritten, is, that ATI "improved the filtering" over its predecessor with the Rage 128 Pro series of cards. I was very curious if I could find the right games to clearly display the improvements that ATI achieved with the new Rage 128 generation and did several tests, just to discover.. nothing. Well almost nothing.. and that's even more interesting.

                        The first very interesting fact is, that a Rage 128 GL can do much better texture filtering than the Rage 128 Pro does and this is solely achieved by picking the right driver. And this would be version 4.11.6216, which does not support the 128 Pro. Out of all the driver versions, this one is hands down the best when it comes to texture filtering, by a long shot. Coming up are some screenshots / crops for comparison between the 4.11.6216 driver and later versions.

                        The left part of the pictures always comes from a Rage 128 Pro with driver version 4.12.6292 from October 2000, which is my go-to driver for all Rage 128 cards. I covered all Rage 128 driver versions in detail in a separate article, for now just be informed that this driver represents the image quality you will get out of any usable (=late-ish) driver on all Rage 128 series of cards.

                        The right part of the picture always comes from a Rage 128 GL rendering with the aforementioned driver version 4.11.6216, released in December 1999.

                        Please click on the images and use the magnification tool to compare the rendering outputs.

                        The uper scene is taken from Unreal (1998). Please notice the keyboards on the right, where the difference can be seen best. Additionaly, the lower right screen on the left side of the picture looks like it is diagonally cut on the Rage 128 Pro. Half of it is filtered correctly, the other half is just mush.

                        The lower crop is taken from the helicopter scene of 3DMark 2000. The ground textures receive much better filtering on the Rage 128 GL to the right, using the older driver.

                         

                        The next comparison originates from the adventure scene of 3DMark 2000. This is one of the most striking examples, as we can clearly see a loss of detail looking at the front of the house. The difference in filtering is also perfectly visible in the castle tower's wall textures.

                        And this is where it gets interesting again. Like in all examples before, we have the sample of the Rage 128 Pro on the left side and the Rage 128 GL on the right. But the difference is, that the Rage 128 Pro all of a sudden does the better filtering in this theoretical test (3DMark99max Quality Test Bilinear Filtering). And to answer the unasked question right away: Yes, the Rage 128 Pro produces the same output as seen here with every compatible driver. And yes, so does the Rage 128 GL. To me this is a clear cheat to make the filtering of the Rage 128 Pro look better in a widely used theoretical test back in the day.

                        It is important to understand that both generations of Rage 128 cards produce the exact same level of filtering quality when using the same driver. The 3DMark99max Quality Test (as well as the identical Quality Test of 3DMark 2000) is the only time the filtering of the Pro looks better. And I did a lot more testing than displayed above, for example using the very useful D3D AF-Tester utility, which visualises not only the quality of anisotropic filtering, but also other filtering such as (bi-)linear filtering and point sampling.

                        The thing is: The rendered images in the AD-Tester look exactly the same on both generations of cards when using the same driver version. On top of that, the applied filtering is also absolutely identical for all driver versions supporting both cards. So I won't bother you with lots of images, that all look the same. However, the one thing I noticed is, that the point sampling seems to have changed from driver version 4.12.6277 and earlier to 4.12.6292 and up. But the difference is so subtle, that it is only noticeable if you look at the images alternating in GIF-like style. I also couldn't tell if one or the other looks "better". But for the sake of completion of this chapter, here are the two outputs, version 4.12.6277 to the left and 4.12.6292 to the right:

                        Feature Test: DXTC Texture Compression

                        Welcome to the second part of our mythbusting adventure. The next feature that we'll try to prove missing, is the texture compression DXTC.

                        Even in the earliest (p)reviews of the Rage 128 Pro from October 1999, which is roughly two months before retail availability, DXTC was listed as a supported feature of the new chip. DXTC is a texture compression technology, which is part of the DirectX 6.0 specifications. It can save memory bandwidth for supported hardware when 32-bit rendering and high resolutions are used.

                        Well, as can be seen in the D3D and OpenGL feature tables from earlier, the feature reports of all drivers compatible with the Rage 128 Pro list this feature as unsupported.

                        Now, I think most of you that have used a Rage 128 Pro based card before, have installed the latest reference driver 4.13.7192 and have probably seen a checkbox in the D3D properties, which should enable Compressed texture format. The option to enable this feature in the D3D driver properties appears the first time in driver version 4.13.7078 from March 2001, far more than year after the card's release and around one year after the release of its successor, the Radeon DDR!

                        D3D properties driver 4.13.7192. Sorry for the german screenshot, "Komprimiertes Texturformat" stands for compressed textures.

                        I think we can agree that this is pretty late for a feature that is supposed to be supported from the beginning. That being said, in theory the checkbox could just be missing and the feature had to be requested by the game. So let's get practical.

                        One of the most commonly known older 3D applications to provide visible support for compressed textures in its options, is 3DMark2001SE. In any case, the above-mentioned checkbox in the driver properties definitely has an effect on 3DMark2001SE, as the option to select "Compressed Textures" for a Rage 128 Pro is not provided at all, unless the checkbox is ticked. I also experimented with enabling compressed textures using the Rage 128 Tweaker for the older driver versions that don't have the checkbox in the drivers. This was to see if the options in 3DMark would change, but it didn't make a difference.

                        So, judging by 3DMark2001SE, the feature really (supposedly) came as late as with driver 4.13.7078. But does it have any effect?

                        Well yes, a tiny little. In a standard 3DMark2001SE run, a Rage Fury Pro gains around 4% with compressed textures enabled, an Xpert 2000 Pro gains about the same (even a litte less). I would have imagined that the Xpert card would benefit more, because of its 64-bit narrow memory interface. However, this was not the case.

                        For comparison, an ATI Radeon SDR gains around 9,5% with texture compression enabled. And, by the way - Everest lists DXTC as a supported feature for the Radeon, as opposed to the Rage 128 Pro. Consequently, I had to "collect" some more samples to get a clearer picture.

                        Unfortunately, it is rather hard to find games from that era, that support texture compression and are not based on the Quake 3 engine, which uses OpenGL. After some research, the only Direct3D game I could find was Expendable. This game lets you enable compressed textures regardless of whether your hardware supports it or not. And the results are equally meaningless: The Rage 128 Pro performs identically with or without texture compression enabled, and so does the Radeon (within the margin of error).

                        Running out of ideas, I finally ended up using Quake 3 Arena for one last test. And eventually, I got results that showed some real differences.

                        But before we get to these, let's summarize what we have found out so far.

                        • Compressed textures enabled in 3DMark2001SE provide only minor gains in performance for Rage 128 Pro cards, even for those with very little memory bandwidth, such as the Xpert 2000 Pro
                        • A Radeon SDR benefits more than twice as much from compressed textures enabled in 3DMark2001SE, but the plus in performance is still just under 10%
                        • In Expendable, compressed textures do not improve performance for Rage 128 Pro based cards, the same applies to the Radeon SDR
                        • For the Rage 128 Pro, Everest reports texture compression as unsupported for both Direct3D and OpenGL with any driver version
                        • In contrast, texture compression is reported as supported for both APIs for the Radeon

                        Another aspect worth discussing is, that compressed textures can only be enabled for Direct3D in the Rage 128 Pro driver. I am not shure if this even makes sense - I would expect it to function the same in OpenGL, as we are talking about a feature that is supposed to be supported in hardware. Sure, software (the driver) has to be written to utilize this part of the chip, but then again - an announced (hardware-) feature that makes it only halfway into the driver around one and a half year after the product shelved? How long can it take for a driver team to get a feature to function, that was claimed to be supported from the beginning?

                        Either way, I still owe you the results for my Quake 3 Arena test.

                        • Game Version 1.32
                        • High Details, Trilinear Filtering
                        • 1024x768x32
                        • Compressed textures were en- and disabled by the commands r_ext_compressed_textures 1 and 0 respectively, followed by vid_restart before each benchmark run.
                          Compressed Textures Disabled Compressed Textures Enabled
                        Rage 128 Pro Ultra 32MB 29,1 FPS 29,1 FPS
                        Radeon SDR 32MB 50,4 FPS 69,0 FPS

                        As you can see, the texture compression setting has zero impact on the performance of the Rage 128 Pro, while the Radeon experiences a substantial boost.

                        Given, the absence of hardware support for texture compression is not as undeniably clear as it is for anisotropic filtering, but it's hard to find solid evidence for the absence of an invisible feature, especially when, on top of that, it is so hard to find software for to check its presence.

                        I invite everyone to decide for themselves, but to me this whole thing smells fishy.

                        Image Quality And Futher Features

                        The image quality of all Rage 128 series cards has some minor drawbacks when rendering in 16-bit mode. The good news is that you can always use 32-bit mode for all games supporting it. The performance hit is nowhere near as drastic as experienced with all other competing chips of the same generation, and the image quality is excellent at higher colour depths.

                        Alpha Blending & Dithering

                        The main issue with these cards in 16-bit colour depth is the ugly dithering in alpha blending scenarios. The only situation that I found it really striking (and distracting at the same time), was in the Gothic level of Unreal Tournament. Everytime I got past one of these lamps inside, there were giant, ugly coronas around them. Unfortunately, these don't come across as well on an 800x600 resolution screenshot, especially since the in-game gamma level seems to be much higher than seen in the original screenshot.

                        Consequently, I upscaled the screenshot to 1600x1200 (that is also the way I looked at it on my LCD) and raised the gamma to a level that makes the image look more in line with what I saw on my monitor.

                        These coronas seen in the screenshot are really distracting when you actually move in the game, I hope this picture gives you a good idea of the problem. Please also notice the stipple pattern in the coronas, which is sympthomatic for ATI's dithered alpha blending with the Rage 128 series.

                        Please don't mind the broken HUD on the upper part of the screenshot, it looks fine in-game.

                        The interesting part about this ugly dithering pattern is, that it isn't present in every driver and can be turned off (with the help of external tweaking utilities) in some older driver versions. The difference in rendering can be visualized best in the game Viper Racing, with the smoke from the tyres at the start of the integrated benchmark.

                        We'll have a look at some pictures in a second, but let me quickly list "the history of dithering" in the Rage 128 drivers (in context of the upcoming Viper Racing example):

                        • 4.11.6060 (launch driver Rage 128 GL): no stipple pattern, nice looking smoke (bit of a grainy look)
                        • 4.11.6076 - 4.11.6216: ugly, stippled dithering; dithering can be turned off with an external tweaker, which removes the stipple pattern, makes smoke looking nicer
                        • 4.11.6263 (launch driver Rage 128 Pro): no stipple pattern, best looking smoke of all drivers
                          • From this driver onwards, the setting in the external tweaker had no effect anymore
                        • 4.12.6269 and up: ugly, stippled dithering

                        The following screenshots are ordered from the visuals I consider to be the worst to the best (in my personal opinion). The first image shows the stippled alpha blending for driver versions 4.11.6076 to 4.11.6216, as well as 4.12.6269 and up. The second image shows the same example with dithering disabled for version 4.11.6076 to 4.11.6216. The output seen on the third and fourth images are unique to the respective driver versions listed below the images.

                        4.11.6076 default

                        4.11.6076 dithering disabled

                        4.11.6060 default

                        4.11.6263 default

                        I found it really interesting that the Rage 128 Pro launch driver (4.11.6263) provides the best visual quality, which is in line with what reviewers back in the day said about the Pro having improved in rendering quality over the GL. Little did they know (or did care to check), that this driver enhanced the rendering quality in the same way for the older generation chip..

                        As I said before, with the exception of Unreal Tournament, I cannot say that this issue bothers me much when using the Rage 128 cards. Even on a still picture, the difference is rather subtle in most cases. Furthermore, you can always use 32-bit rendering with newer drivers and games, as well as maybe try the older 4.11.6263 driver for games that lack support for 32-bit colours.

                        If you would like to see some further comparisons, I invite you to check out my article comparing all Rage 128 drivers.

                        Paletted Textures

                        Some older games require a graphics card that supports paletted textures. Notable examples are Final Fantasy VII and VIII; the former will even refuse to start if this feature is not supported. That being said, this was changed later with a patch. Nevertheless, some games don't look right without support for paletted textures, which makes it an important feature for some people.

                        A list of games that require or make use of paletted textures can be found on vogonswiki.com.

                        The latest ATI reference driver (version 4.13.7192) has working support for paletted textures, making it the only driver that makes this feature available on a Rage 128 Pro. For the Rage 128 GL, you can also use all drivers up to and including version 4.11.6216, which is the final driver only supporting first generation Rage 128 cards.

                        There is a checkbox in the D3D properties of all driver versions between the latter and the latest reference driver to enable this feature, however, it still doesn't work in any but the aforementioned driver versions. I took a screenshot of the Final Fantasy VIII settings menu to demonstrate the difference. The image on the left shows paletted textures working, while the image on the right shows the result for all driver versions with broken support for them.

                        Table Fog

                        Another important thing people are interested in on retro graphics cards (even more so than in paletted textures), is the support for table fog. The visual difference when using hardware without support for the latter feature in games that require it to render correctly is definitely more noticeable and distracting than the difference if support for paletted textures is missing.

                        Furthermore, the list of games making use of table fog is a lot more extensive.

                        On the Rage 128 series of cards, support for table fog is kind of a mixed bag. Firstly, table fog is broken on all 4.13.xxxx drivers, which includes the latest reference driver. Secondly, even on drivers that have working support for table fog, it doesn't seem to work with every game.

                        I picked Star Wars - Shadows of the Empire and Thief 2 for the test and while the fog in Thief renders correctly (on all but the 4.13.xxxx drivers), the fog in Shadows of the Empire is completely missing. I also tried 3D Winbench 97, which is a pretty solid software to check for the support of 3D features of vintage 3D accelerators, but the fog wasn't rendered either in the respective test.

                        I could imagine that the support for table fog on the Rage 128 series of cards is done in software, possibly comparable to what nVidia did with their Riva 128 and TNT series of cards, where table fog was emulated using vertex fog. But then again, the Riva 128 (ZX) at least renders the fog correctly in 3D Winbench 97 as well as in Shadows of the Empire, which can be seen in a comparison I did several years ago on VoodooAlert: 3DWB97 & SOTE.

                        In the pictures below you can see the table fog in Thief 2 working correctly on the left with driver version 4.12.6292 and the broken rendering of version 4.13.7192 on the right image. Below these are images of the 3D Winbench 97 table fog test. The image on the left is a reference image and the image on the right shows the rendering output of any Rage 128-based card combined with any driver version.

                        Table fog working with 4.11.xxxx and 4.12.xxxx drivers

                        Broken table fog support in 4.13.xxxx drivers

                        Reference Image

                        Missing table fog on all Rage 128 cards and drivers

                        If you'd like to see some more samples, including Shadows of the Empire, again, you can check out my article comparing all Rage 128 drivers.

                        All These Models

                        At this point, I think we have covered all the important aspects that have to do with everything but the actual hardware, the cards themselves. This brings us to the final part of this overview, which will focus on all the different cards hiding behind ATI's product numbers, as well the performance of each card.

                        The sheer number of different models (and revisions of those) of basically four different chips across two generations is unbelievable and hardly rivaled even to this day (certainly not at the time these cards were released). Therefore, let me kick this off with some guidelines on how to tell some of these apart or identify certain details with just a quick look.

                        Some General Help With Identification

                        If you are still a rookie on Rage 128 territory, you will have a hard time identifying Rage 128 cards on small pictures, bad angles or just a quick look. But don't worry, here are some easy tips that may prove useful when you can't find or see the ATI product number on the PCB.

                        Generation 1 vs. 2 / Rage 128 (GL/VR) vs. Rage 128 Pro (VR)

                        This one is fairly easy: Rage 128 based cards are AGP 2x (3,3V) only, Rage 128 Pro cards also support AGP 4x and 1,5V, a quick look on the AGP connector is enough to identify the generation:

                        Generation 1 - AGP2X Connector (109-66500-11)

                        Generation 2 - AGP4X Connector (109-65700-20)

                        The second thing to tell the two generations apart, is the size of the heatsink. The Pro cards have a huge black heatsink, 1st gen Rage 128 cards either have a small all black heatsink, or a small black heatsink with Rage 128 written on it.

                        Generation 1 - Small Black Heatsink (109-51900-01)

                        Generation 1 - Small Heatsink with Logo (109-66500-11)

                        Generation 2 - Big Black Heatsink (109-65700-20)

                        But attention: The second method is not completely bulletproof, as there is a 109-51900-31 variant, which is a Rage 128 GL based card, but has a big black heatsink installed. This is the only exception of the rule that I know of, but if you want to be shure, pick method one.

                        Spot Early Rage 128 GL/VR Cards Easily

                        The first and most reliable thing to look out for, is the style of the heatsink (see pictures above). If there is a Rage 128 logo on it, the card is a later one. If the card has a small all black heatsink installed, it is an early card. Equally reliable (but depending on the picture maybe hard to spot) is the presence of ceramic capacitors on the backside of the chip. All early cards have those installed to ensure stable operation, later cards didn't need them anymore.

                        Early Card - MLCCs installed (109-51900-01)

                        Later Card - no MLCCs present (109-50500-11)

                        Two other indicators which are found on all early cards (but can also be present on later cards, if the OEM did some "leftover recycling") are a black VGA connector and a big oscillator chip for clock generation. Later cards have a small clock generator and a blue VGA connector:

                        Early Card With Black VGA Connector And Oscillator (lower right corner) (109-51900-01)

                        Later Card With Blue VGA Connector And Regular Clock Generator (109-50500-11)

                        Easily Avoid Cards With 64-Bit Memory Interface

                        • All L-shaped cards have a 64-Bit memory interface, no matter if they are Rage 128 oder Rage 128 Pro based
                        • All cards with "missing" memory chips have a 64 Bit memory interface, as long as they have SDR memory.
                          Cards with SG-RAM still have a 128-Bit memory interface with only four out of eight memory chips installed.
                        • To put it the other way around: All cards with SG-RAM have a 128-Bit memory interface.
                        • Unfortunately, the "missing memory chips" identification isn't bulletproof in the opposite direction (at least for the first generation of cards):
                          • Rage 128 VR cards always have a 64-Bit memory interface, although they don't have "missing" memory chips
                          • There are also Xpert cards that are not L-shaped and do not have "missing" memory chips, which can be misleading at first glance, for example:

                        ATI Xpert 99 (109-52000-00) (Rage 128 VR)

                        ATI Xpert 2000 (109-68100-01) (Rage 128 GL)

                        Rage 128 (GL/VR/Pro/Pro VR): All Models Listed

                        First things first: All the information in the upcoming list, as well as the even more detailed information on each card/model on its respective detail page, has been collected by me from each card itself. No technical information in my lists is taken from third-party sources, which could contain inaccurate information. All specs are supported by pictures and/or screenshots.

                        In the case of community cards that are included here, I try to get the same depth of information that I collect and only confirmed information gets included in the list and on the detail pages.

                        One click on the part number will take you to the detail page of each card.

                        Part Number Model Chip MemSize MemType MemBits Core Mem
                        109-52000-00 Xpert 99 Rage 128 VR 8 MB SDR 64 79 125
                        109-50500-00 Rage Magnum Rage 128 GL 16 MB SGR 128 90 90
                        109-50500-11 Rage Magnum Rage 128 GL 16 MB SGR 128 90 90
                        109-51800-01 Xpert 128 Rage 128 GL 16 MB SDR 128 103 103
                        109-51900-31 Rage 128 OEM Rage 128 GL 16 MB SDR 64 80 125
                        109-52900-02 All-in-Wonder 128 Rage 128 GL 16 MB SDR 128 90 90
                        109-51900-01 Rage Fury Rage 128 GL 32 MB SDR 128 90 90
                        109-51900-01 Rage Fury (Retail) Rage 128 GL 32 MB SDR 128 103 103
                        109-52100-01 Xpert 2000 Rage 128 VR 32 MB SDR 64 80 80
                        109-53400-11 All-in-Wonder 128 Rage 128 GL 32 MB SDR 128 92 92
                        109-61300-00 Rage Fury Rage 128 GL 32 MB SGR 128 90 90
                        109-66500-00 Xpert 2000 Rage 128 GL 32 MB SDR 64 80 105
                        109-66500-11 Xpert 2000 Rage 128 GL 32 MB SDR 64 80 120
                        109-68100-01 Xpert 2000 Rage 128 GL 32 MB SDR 64 79 79
                        109-60600-10 Rage 128 Pro OEM Rage 128 Pro 16 MB SDR 64 120 120
                        109-63100-10 Rage Fury Pro OEM Rage 128 Pro 16 MB SGR 128 118 140
                        109-63200-01 Rage Fury Pro OEM Rage 128 Pro 16 MB SDR 64 120 120
                        109-70400-01 All-in-Wonder 128 Pro PCI Rage 128 Pro 16 MB SDR 64 ? ?
                        109-73100-01 Rage 128 Pro Ultra Rage 128 Pro2 16 MB SDR 64 130 130
                        109-63100-10 Rage Fury Pro Rage 128 Pro 32 MB SGR 128 118 140
                        109-63100-11 Rage Fury Pro OEM (DVI) Rage 128 Pro 32 MB SGR 128 120 120
                        109-68200-00 Rage Fury Pro Rage 128 Pro 32 MB SGR 128 ? ?
                        109-63200-01 Rage Fury Pro OEM Rage 128 Pro 32 MB SDR 128 120 120
                        109-63200-01 Rage Fury Pro (Retail) Rage 128 Pro 32 MB SDR 128 118 140
                        109-65600-01 All-In-Wonder 128 PRO Rage 128 Pro 32 MB SDR 128 118 140
                        109-65700-20 Xpert 2000 Pro Rage 128 ProVR 32 MB SDR 64 120 120
                        109-73100-02 Rage 128 Pro Ultra Rage 128 Pro2 32 MB SDR 64 133 133
                        109-78200-00 Rage 128 Pro Ultra Rage 128 Pro2 32 MB SDR 64 120 120

                        Performance Comparison

                        As promised before, we'll now come to a performance comparison of all ATI Rage 128 GL/VR/Pro/Pro2 models. This is internet's first and will hopefully help many people interested in ATI's pre-Radeon generation of cards who don't know which card to look for. In any case, these are the systems I used for the benchmarks:

                        Test System #1 (AMD K6, Super Socket 7 Platform)

                        Retromachine: RHR Benchmarking System Super Socket 7
                        This system comes to use when it fits the hardware being tested and represents the lower end of a 3D retrogaming PC. It is fast enough for early 3D titles, but is also very useful for exhibiting CPU bottlenecks when compared to faster systems with the same graphics card. The system itself is not exactly optimized (on purpose), so there are definitely faster systems with a K6-2 450 out there. It should represent a bread-and-butter K6-2 450 system from back in the day pretty accurately. Hardware: Gigabyte GA-5AX Rev. 4.1 (ALi Aladdin V) AMD K6-2 450MHz (4,5×100) 128MB PC133 CL2... Continue reading

                        Test System #2 (Intel Pentium III, Slot 1 Platform)

                        Retromachine: RHR Benchmarking System Slot 1
                        This system comes to use when it fits the hardware being tested and represents (in its “default” configuration) a capable 3D retrogaming PC, without being top tier. It doesn’t bottleneck most DX6.1 GPUs (as long as the resolution ist 1024×768 or higher) and is also still suitable for DX7 graphics cards. Another benefit of the Slot 1 platform is the possibility to scale CPU performance from a Celeron Covington clocked 133 MHz all the way to a 1400 MHz Tualeron. But its default config is as follows: Hardware: Asus P3B-F Rev. 1.04 (Intel i440BX) Intel Pentium !!! 850 MHz (8,5×100)... Continue reading

                        Test System #3 (AMD Athlon, Socket A Platform)

                        Retromachine: RHR Benchmarking System Socket A Classic
                        This system comes to use when it fits the hardware being tested and represents a top-end 3D retrogaming PC for all graphics cards up to DX7 level. Hardware: EPoX 8KTA3 (VIA KT133A) AMD Athlon Thunderbird 1400C (10,5×133) Infineon 256MB PC133 CL2 2-2-5 Maxtor DiamondMax Plus D740X 20GB Creative Soundblaster 128 PCI 3Com Fast Etherlink XL PCI Software: Windows 98SE Unofficial Servicepack 2.1d DirectX 8.1 VIA 4in1 4.56 Creative -some version from my fileserver- To make sure that all results are correct and stay comparable even if further cards are added to a comparison that this system was used for, I... Continue reading

                        All benchmarks were done using driver version 4.12.6292, as this is the best performing driver.

                        Descent³: secret.dm2 timedemo

                        Descent³, released in mid-1999, featured the newly developed Fusion Engine, which provided broad hardware support. It supported all the most common 3D APIs at the time, such as Direct3D, OpenGL and 3dfx Glide, as well as EAX and A3D for 3D audio hardware. The engine also had a state-of-the-art feature set with regard to 3D effects. Descent³ runs relatively decent (no pun intended) on low-spec systems, but also scales very well with both more CPU and GPU horsepower.

                        API: Direct3D | Version 1.4 | Settings used:

                        Note: VSync was turned off in the driver, but had to be turned on ingame, otherwise it would stutter heavily.

                        AMD K6-2 450 MHz

                        The CPU is too weak to reveal any significant differences between the various graphics cards, especially at 800x600. However, even with the K6, at 1024x768, the performance of cards with a 64-bit memory interface begins to drop significantly, particularly for first-generation VR cards.

                        I also wasn't able to complete the benchmark at 1024x768 with the 8 MB 109-52000-00 VR card, as it would freeze every time at that resolution.

                        Pentium III 850 MHz

                        Things look a lot different with the Pentium III system and we are able to see a couple of interesting things. Starting from the top, it is interesting to see that Rage 128 Pro cards with SGR memory have have an advantage of around 10% at the higher resolution over their similarly clocked SDR counterparts. It is also obvious, that 16 MB of memory is absolutely sufficient for Descent³.

                        Moving towards the middle, we have the Rage 128 GL top dogs beat all Rage 128 Pro cards with a 64-bit memory interface in both resolutions. These two cards (109-51800-01 and 109-51900-01 Retail) clock at 103/103 MHz and are significantly faster than all other Rage 128 GL cards with a clock speed of 90/90 MHz. The latter perform about at the level of Rage 128 Pro cards with a 64-bit memory interface, with a slight edge at the 1024x768 resolution.

                        First-gen cards with a 64-bit memory interface only achieve acceptable framerates at the 800x600 resolution. The 8 MB VR card refused to complete the benchmark in both resolutions on the Intel system.

                        Athlon 1400C

                        The Athlon system provides another nice performance boost for the more powerful cards. At the top we find the highest clocked Rage 128 "Ultra" card (109-73100-02; 133/133 MHz), which clearly dominates at the 800x600 resolution. At 1024x768, however, all Rage 128 Pro cards with SGR memory are still a little bit faster.

                        The rest of the field looks similar to what we saw with the Pentium III system. The overall performance gain of the cards decreases towards the lower end of the chart.

                        Half-Life Opposing Force: Blast From The Past timedemo

                        Opposing Force was the first expansion pack for Half-Life and got released roughly a year after the original game, in November 1999. I used the "blast" timedemo, that a member of voodooalert.de recorded. He also put a ready-to-bench package including the demo version of the game together, which can be found on archive.org.
                        The original thread with several other results can be found here.

                        API: OpenGL | Demo Version 1.0.1.2 | Settings used:

                        - none configurable / default

                        AMD K6-2 450 MHz

                        The results for the AMD K6 test system are even closer in this bench, than in the Descent³ benchmark before. The Opposing Forces timedemo is a lot more CPU limited, that's why we see next to no difference between the two resolutions tested for almost every card. The only card that experiences a heavy FPS drop at the 1024x768 resolution, is the 8 MB 109-52000-00, which seems to run out of video memory above 800x600.

                        Pentium III 850 MHz

                        The 850 MHz Pentium III is already powerful enough to produce a proper scaling.

                        We have the 109-73200-02 leading the pack in both resolutions, followed by the 128-bit Rage 128 Pro cards, which all perform pretty similar. This benchmark doesn't seem to benefit from SGRAM like Descent³ does and memory bandwidth in general doesn't seem to help that much here, as the 118/140 MHz clocked cards (109-63200-01 ViVo and 109-63100-10) perform relatively identical to all other Rage 128 Pro cards with 128-bit memory, which clock at 120/120 MHz.

                        Another indicator is the 109-73100-01 leading the midfield, which only has 16 MB 64-bit memory, but is clocked at 130/130 MHz. The higher core clock seems to do more here.

                        The rest of the 64-bit Rage 128 Pro cards get slightly outperformed by the fastest Rage 128 GL cards again, while the former are a good bit ahead of the remaining Rage 128 GL cards, at least at the 800x600 resolution. At 1024x768 they perform about the same.

                        The Rage 128 VR cards are too slow to hit the 30 FPS mark at 1024x768 and populate the lower end of the chart.

                        Athlon 1400C

                        With the Athlon system we get pretty much the same picture, but the 109-73200-02 manages to pull ahead even more at the 800x600 resolution. At 1024x768, there is no futher scaling across the board. Even the fastest card gains less than 0.5 FPS.

                        MDK2: integrated benchmark

                        MDK2, released in May 2000, is the newest title among the selected benchmarks. It already supports Hardware T&L and has the highest demand on video memory out of all tests.

                        API: OpenGL | Version 1.002 | Settings used:

                        AMD K6-2 450 MHz

                        While we already have a lot going on in the MDK2 bench with the AMD K6, we can still see a heavy CPU limit, especially at the lower 800x600 resolution. All cards with 32 MB of memory and a 128-bit memory interface achieve around 25-26 FPS in both color depths at that resolution. On top of that, Rage 128 Pro cards with the aforementioned memory configuration can even keep this framerate at 1024x768 in both color depths.

                        But another thing also becomes clear, even with the underpowered K6-2 450: cards with less than 32 MB memory experience a significant drop in performance at 1024x768x32 in MDK2 at the highest settings.

                        The slowest Rage 128 VR cards at the bottom of the chart do not need a faster CPU to reveal how underpowered they are for the task. These cards are already GPU limited and achieve only around 20 FPS in this system at 800x600x16.

                        The 8 MB 109-52000-00 completely runs out of memory at that resolution and can't even run higher resolutions and/or color depths.

                        Pentium III 850 MHz

                        The Pentium III introduces a proper scaling and brings a massive performance boost for the faster cards, even at 1024x768 in 32-bit color depth.

                        We have both Rage Fury Pro cards and the All-in-Wonder 128 Pro in first, second and third place, all clocking at 118/140 MHz. The SGR card (109-63100-10) shows advantages in both color depths at the higher resolution over the other two cards, which are equipped with SDR memory.

                        MDK2 is another example of how SG-RAM can help with higher resolutions, as the lower-clocked Fury Pro cards (120/120 MHz) in fifth and sixth place show a similar sorting.

                        Memory bandwidth in general seems to be a thing in this game. The Rage 128 Ultra (109-73100-02) (133/133 MHz) in fourth place performs about the same as the SDR Fury Pro cards in front of it, which have ~5% higher memory clocks but clock ~12% lower on the chip.

                        Another indicator for MDK2's appetite for memory performance are the numbers of all cards with a 64-bit memory interface. These are all on the bottom end of the list, no matter if they carry a first- (128 GL/VR) or second-gen (128 Pro) chip. The exception of the rule is the 109-73100-01, which benefits of its rather high clocks (130/130 MHz) and performs around the level of a 16 MB Rage 128 GL card with a 128-bit memory interface.

                        Before we move on to the results of the Athlon system, I want to point out another observation which carries through all the benchmarks, but becomes very clear in MDK2:

                        Rage 128 GL cards with SG-RAM are not able to benefit from this type of memory. Both the 16 MB and 32 MB version are the slowest Rage 128 GL cards with a 128-bit memory interface and perform about 10% lower then their SDR counterparts across all resolutions and color depths.

                        Athlon 1400C

                        The Athlon 1400C primarily brings slightly higher numbers at 800x600x16 for the first few cards. Don't let the sorting fool you, the 109-7100-02 is ranked first because of its performance in the 800x600x16 test. However, the fastest card at 1024x768 in both color depths is still the 32MB 109-63200-10.

                        The numbers in higher resolutions and color depths for all cards are about the same (sometimes even completely identical), so we have a healthy GPU limit here.

                        However, there is still another card I need to highlight again, and this is the "retail clocked" (103/103 MHz) 109-51900-01. This Rage 128 GL card manages to outperform the 16 MB variant of the 109-63100-10 (128-bit SGR Rage 128 Pro) and ranks seventh. A good performance of the original Fury!

                        Quake 2: demo1.dm2 timedemo

                        No 3D benchmarking suite covering 3D graphics up to the millennium is complete without Quake 2. While this game is the oldest of the selected tests, it still scales beautifully with stronger hardware. The demo1.dm2 timedemo used here is also a lot less CPU dependant than the famous crusher.dm2 timedemo.

                        API: OpenGL | Version 3.20 | Settings used:

                        AMD K6-2 450 MHz

                        Quake 2 is no unsolvable task for all cards, especially at the 800x600 resolution. Most cards are even able to achieve over 30 FPS at 1024x768 in 16-bit color depth. However, 32-bit at that resolution is a bit too much for first-gen Rage 128 cards with a 64-bit memory interface, some of them even fall below 20 FPS.

                        The 8MB 109-52000-00 runs out of video memory in this scenario and can't run the timedemo at 1024x768x32.

                        Pentium III 850 MHz

                        The plus in CPU performance brings something very interesting to light. Across the board, we see only a very tiny performance hit in 32-bit for cards with a 128-bit memory interface, even at the higher 1024x768 resolution. This is something contemporary reviewers highlighted for both the original Fury as well as the Fury Pro.

                        While the performance decrease in newer, more complex games such as MDK2 is still rather moderate compared to their respective competitors, the performance hit in earlier 3D titles with 32-bit support is only 5% or less. And this is not limited to Quake 2, as we will see later.

                        Be that as it may, Quake 2 doesn't care too much for the type of memory or anything - raw power wins. The 109-73100-02 leads the pack, followed by all Fury Pros with a 128-bit memory interface.

                        Memory size and and clockspeed are next to meaningless among the latter, only the SGR versions tend to have a minor advantage of less then 5% at the 1024x768 resolution.

                        The Rage 128 Ultra 64-bit (109-73100-01) shows about the performance of a Rage 128 GL with 103/103 MHz clocks, with the latter being faster in 32-bit color depth in both resolutions.

                        The midfield consists of all Rage 128 Pro cards with a 64-bit memory interface, as well as 128-bit Rage 128 GL cards operating at 90/90 MHz. The Xpert128 (109-5800-01) sits between these two groups and is faster than all the 64-bit Rage 128 Pro models in the midfield at all resolutions but 800x600x16, thanks to its higher clockspeeds of 103/103 MHz.

                        Athlon 1400C

                        The Athlon system only adds minimal performance, and even then, only at lower resolutions and color depths for the fastest bunch of cards. Even the 109-73100-02 in first place only gains 0.4 FPS in 1024x768x32, falling slightly short of the 60 FPS mark. A fast Pentium III seems to be more than sufficient for any Rage 128 card in Quake II.

                        Unreal: castle flyby

                        Unreal was released in mid-1998 and raised the bar for visual fidelity in 3D games to another level. Upon release, it only supported 3dfx Glide and PowerSQL for 3D acceleration, but patches also brought support for OpenGL, Direct3D and even S3 MeTaL. This makes it an ideal benchmark for hardware of that era.

                        API: Direct3D | Version 2.26 | Settings used:

                        • gamestart with -firstrun parameter
                        • this ensures that all advanced settings are on default on every benchmark run
                        • no further settings are changed

                        AMD K6-2 450 MHz

                        The K6-2 system is definitely underpowered for Unreal. We have several erratic results, were some cards are faster in 32-bit than they are in 16-bit, cards with a 64-bit memory interface placed towards the top and so on.

                        One thing we can already see is that the 8 MB 109-52000-00 obviously has insufficient video memory, running out completely at 1024x768 in 32-bit. While this model can be one of the fastest VR cards if the size of the VRAM is not the limiting factor, it falls to last place at all settings.

                        Pentium III 850 MHz

                        In the faster Pentium III system, we have all Rage 128 Pro cards with a 128-bit memory interface on top. The 16 MB 128-bit 109-63100-10 falls back a few FPS at the 1024x768x32 resolution, so Unreal seems to need more than 16 megs of memory at that setting.

                        The 109-71300-02 in first place can't pull away from the other Rage 128 Pro cards in 800x600, despite its higher clockspeeds. At this resolution, we are still CPU limited in both color depths.

                        Unreal seems to be another game, where cards with SGR memory are faster at 1024x768, in 800x600 performance is about the same as seen on SDR models. But this is only true for Rage 128 Pro cards, the SGR-equipped Rage 128 GL models 109-50500-00 and 109-61300-00 are once more a tad slower than their SDR counterparts in every setting.

                        The first-gen top-dog 109-51900-01 'retail' model shows another strong performance, ranking behind the 128-bit Pro cards and providing over 30 FPS at all settings. Although the high-clocked 64-bit Rage 128 Ultra (109-73100-01) is even a bit ahead of it due to its slightly better performance at 16-bit, it falls behind the Rage Fury at the more demanding 32-bit setting.

                        Following closely is the Xpert128 (109-51800-01), which ranks ahead of the midfield consisting of all the other 128-bit Rage 128 GL and 64-bit Rage 128 Pro models.

                        At the end of the test field are the 64-bit first-gen cards, with the slowest among them barely breaking the 30 FPS barrier at 800x600x16.

                        Athlon 1400C

                        We get the same picture with the Athlon system, only with increased performance for pretty much every card, excluding the four 32 MB 64-bit first-gen cards at the end of the chart. The 8 MB 109-52000-00 is able to exploit the plus in CPU performance more effectively than these models, allowing it to move away from the bottom of the list.

                        Simultaneously, the Rage 128 Ultra in first place shows an impressive performance increase, particularly at the 800x600 resolution. This allows it to pull ahead more clearly from the remaining Rage 128 Pro cards.

                        Performance Ranking

                        Up next are the complete performance rankings for each resolution and colour depth, which I will leave uncommented for you to explore. These numbers were collected on yet another system, which I use for global comparisons between different graphics cards:

                        Retromachine: Official RHR Benchmarking System for DX6-DX8 Suites
                        This system is solely used for the global benchmark comparisons on this site. It has enough CPU performance for the fastest DX8(.1) cards out there and its VIA KT333 chipset ensures compatibility to older graphics cards, that can only handle 3,3V AGP. Hardware: EPoX 8KTA3+ (VIA KT333) AMD Athlon XP-m (Barton) @ 2133 MHz (16×133) @ 1,575V 512MB Corsair CMX512-3200XLPRO v1.1 @ DDR266 CL2 2-2-5 Western Digital Caviar WD800BB 3Com Fast Etherlink XL PCI Software: Windows 98SE Unofficial Servicepack 2.1d DirectX 8.1 VIA 4in1 4.56 VIA AC’97 Audio 4.05.00.1130 To make sure that all results are correct and stay comparable... Continue reading

                        The results for the most important models of the following charts also went into my global performance index, where these can be compared to competing cards:

                        RHR Benchmark Suite 1997-2000, DirectX 6+ Performance Index
                        Introduction This post will be continously updated, as all graphics cards that I’ll write about _and_ fit the criteria for this suite, will undergo the tests of my RHR DX6+ suite and will be included in the charts. If you are interested in further information on the suite, I invite you to read the descriptions provided below the charts. They contain details on the games and settings used, the driver versions used for all cards, etc. If you are interested in further information on a graphics card used in the benchmarks, you can search for it in the box on... Continue reading

                        800x600x16

                        1024x768x16

                        800x600x32

                        1024x768x32

                        And that's it! For me, this is the end of the biggest project I ever did (and I had a few 😉 ). I actively worked on this article (and everything that emerged out of this, like the GPU Database and the idea for a global performance index) for over six months this year in my spare time and I am happy that it is finally complete.

                        I want to thank several members of the voodooalert.de and hardwareluxx.de forums for their help with testing, research and aquiring missing cards. You know who you are 🙂

                        I also want to thank you, if you made it through the whole thing! While this project was very exciting for myself (otherwise I wouldn't have done it), I am always happy for everyone else who is also interested in what I am doing.

                        So - until next time 😉

                         - Tweakstone

                        Tagged  , , .
                        Bookmark the permalink.

                        Leave a Reply

                        Your email address will not be published. Required fields are marked *