Benchmark – CPU Scaling 3dfx Banshee on Slot1

Introduction


This is a test I wanted to do for myself for a pretty long time. Without any testing, my answer on how fast a CPU should be so that a Banshee isn't held back would have been "400-500 MHz depending on the card".

 

To finally answer this question thoroughly, I equipped my Slot1 testmachine with seven different Slot1 Pentium II/III CPUs and picked nine Banshees from my collection which represent all combinations of core/mem clocks for SDR- and SGR- Banshee models that I am aware of. This should give us an excellent impression, how far these cards scale and how much of an impact on scaling the core/mem clocks and the memory type installed on the cards have.


The Test System


The testsystem is the following:

 

  • Mainboard:
  • RAM:
  • NIC:
  • Audio:
  • HDD:
Asus P3B-F Rev. 1.04 (Intel i440BX)
Infineon 256MB PC133 CL3 @ PC100 CL2 2-2-X
3Com Fast Etherlink XL PCI
Creative Soundblaster 128 PCI
Western Digital WD400BB (ATA100 / 40GB / 7.200RPM)

For the OS, I used Windows 98 Second Edition with DirectX 8.1 as well as an inofficial Service Pack installed. For each card, I restored a pre-configured system-image before testing.

List Of CPUs Used


 

  • Pentium II 300
  • Pentium II 350
  • Pentium II 450
  • Pentium III 550*
  • Pentium III 600
  • Pentium III 700*
  • Pentium III 850*
Klamath, 66 MHz FSB, 512 KB cache on cardridge, half speed
Deschutes, 100 MHz FSB, 512 KB cache on cardridge, half speed
Deschutes, 100 MHz FSB, 512 KB cache on cardridge, half speed
Coppermine, 100 MHz FSB, 256 KB cache on-die, full speed
Katmai, 100 MHz FSB, 512 KB cache on cardridge, half speed
Coppermine, 100 MHz FSB, 256 KB cache on-die, full speed
Coppermine, 100 MHz FSB, 256 KB cache on-die, full speed

 







* all Coppermine-core CPUs are Socket 370 chips, which were adapted via a MSI Slotket card.

List And Specs Of Banshee Cards Used


Card Core / Mem Mem Type Mem Size
Gigabyte GA-630 110 / 130 MHz SG-RAM 16 MB
Diamond Monster Fusion 105 / 120 MHz SG-RAM 16 MB
Asus AGP-V3200/16M Rev. 2.0 100 / 100 MHz SG-RAM 16 MB
DCS Provideo 3DeMON PV810 105 / 105 MHz SG-RAM 16 MB
DCS Provideo 3DeMON PV810 105 / 105 MHz SD-RAM 16 MB
Miro HiScore Pro 100 / 100 MHz SD-RAM 16 MB
Asus AGP-V3200/8M 100 / 100 MHz SG-RAM 8 MB
Skywell Magic Twinpower 100 / 110 MHz SG-RAM 16 MB
Britek Viewtop Vulcan B 100 / 115 MHz SG-RAM 16 MB

 

The Benchmarks


All benchmarks were made using the latest official 3dfx Banshee driver (1.04.00) for Win9x with the following driver settings (identical for D3D and OpenGL / Glide):

 

  • VSync: off
  • Filter: 4x1 (linear)
  • LOD: +1
  • Alpha-blending: sharper

 

 

These settings were applied via VControl and are my overall recommendation for daily use. The benchmark suite includes 3DMark99 for the D3D API, Unreal Tournament for 3dfx Glide, as well as Quake 2 and Quake 3 Arena, representing an early and a more modern OpenGL title. For both of the OpenGL games, the 3dfx MiniGL driver 1.49 was used.

 

 

All benchmarks were run in the 800x600 resolution. In my personal opinion, this is the best resolution to run games on a Banshee card. Performance is always usable in period-correct games and furthermore, you get integer-correct scaling, if you are using an older 21" TFT with a 1600x1200 native resolution. For those who don't have enough room for a CRT display, these can be great alternatives for retro-gaming.



3DMark99 MAX


This D3D benchmark was released in March of 1999 and is a perfect tool to measure D3D performance of old cards, who do not have a TnL engine. Newer 3DMarks punish cards heavily, that do not support certain hardware features. This edition scales well with raw rendering performance.

The first thing that becomes obvious, is, that 3DMark99 MAX requires at least a 550 / 600 MHz CPU for a fast second-gen 3D accelerator. Even the slowest Banshee (which is the Miro HiScore Pro, representing all SD-RAM equipped models which clock at 100 / 100 MHz) shows a linear scaling all the way up to the 550 MHz Coppermine CPU.

 

 

The fastest Banshees (GA-630 and Monster Fusion) show a notable jump in the 3DMark score coming from the 450 MHz Pentium II. These two cards even scale up to 700 MHz in this test, performance for most other cards remains largely unchanged beyond 550 MHz. Some cards like the Vulcan B with its fast 115 MHz SGR memory gain a few points with faster CPUs, but that's no such thing as scaling.

 

 

The DCS card with SDR memory and 105 / 105 clocks performs about identical as SGR cards clocking at 100 / 100 MHz, such as the two Asus Banshees.

8 MB of memory seems to be no disadvantage in this test, both Asus cards achieve nearly identical scores.



Unreal Tournament: benchmark.dem timedemo


Released in late 1999, Unreal Tournament transformed the Unreal Engine into one of the defining multiplayer shooters of its era. Supporting Glide, Direct3D and OpenGL, it became a popular benchmark for late-1990s and early-2000s hardware. While the game is notoriously known for it's hunger for CPU power, the benchmark.dem timedemo is pretty GPU-limited and can be seen as a worst-case scenario benchmark for UT99.

 

Version 4.51 | Settings used:

Much as the 3DMark99 MAX, UT's benchmark.dem timedemo also suggests the Pentium III 550 MHz as the sweet spot for most cards. Even the GA-630 gains less than 1 FPS when replacing this CPU with a 850 MHz variant. The biggest scaling happens up to 450 MHz. All cards except the Diamond Monster Fusion and the Gigabyte GA-630 achieve less than 1 additional FPS, when replacing the Pentium II 450 MHz with a 550 MHz Coppermine.

 

 

Furthermore, this graph clearly shows that overclocking both the chip and the memory has a significant impact on performance. The trio, consisting of the Skywell model clocking at 100 / 110 MHz and the two DCS cards clocking at 105 / 105 MHz, perform about the same and are already about 2 FPS faster on average than all the cards clocking at 100 / 100 MHz.

 

 

Moreover, the Vulcan B with 100 MHz on the chip but memory clocking at 115 MHz is even faster than these cards, which indicates that already a stock-clocked chip can benefit from the increased memory bandwidth. As the fastest of the bunch, the Diamond and Gigabyte models demonstrate significantly greater performance than all the others due to their unrivalled clock rates.

 

 

The 8 MB Asus card, on the bottom of the graph, struggles in this game. Its average FPS shown in this graph do not tell the whole story - the biggest problem are the minimum FPS, which frequently dip down into the single digits. For cards with only 8 MB of memory, the details should be reduced to smooth out the performance.



Quake 2: demo1.dm2 timedemo


No 3D benchmarking suite covering 3D graphics up to the millennium is complete without Quake 2. Its engine was frequently used in many popular games of the late nineties and reveals the performance in games using a more primitive 3D engine. The demo1.dm2 timedemo used here is a lot less CPU dependant than the famous crusher.dm2 timedemo, so that the performance is limited by the graphics card.

 

Version 3.20 | Settings used:

In Quake 2, the CPU used doesn't matter for a Banshee card. Even the venerable 300 MHz Klamath is more then fast enough for even the fastest Banshee and doesn't take away much of its maximum performance.

 

8 MB of memory is suffient and memory bandwidth doesn't impact performance as much as increased clockspeed of the chip, as shown by the DCS cards which both have 5 MHz overclock on both chip and memory. The SDR model matches the performance of the Skywell card, which has +10 MHz on the (SGR) memory, the DCS card with SGR memory even matches the performance of the Vulcan B, which runs at a memory clockspeed of 115 MHz but has no "overclock" on the chip.



Quake 3 Arena: four timedemo

Released just a few days after UT99 in early December 1999, Quake III Arena introduced id Software's new id Tech 3 engine, which quickly became one of the most influential 3D engines of its era. A major step forward from the Quake 2 engine, it brought a much more advanced rendering pipeline with features such as curved surfaces, higher-detail environments and improved lighting effects. It scales exceptionally well across a wide range of hardware, making Quake III an ideal benchmark for systems from the turn of the millennium.

 

Version 1.32 | Settings used:

Quake 3 Arena shows a relatively similar scaling pattern as 3DMark99 MAX. Something that strikes the eye even more than in the aforementiond benchmark, is, that the 600 MHz Katmai performs worse than the 550 MHz Coppermine Pentium III. This serves as a reminder that architectural improvements can outweigh a slight advantage in clock-speed. The more advanced a 3D engine is, the more it benefits from a fast L2 cache. This is one of the reasons why both Intel and AMD replaced the external L2 cache designs of early Pentium III and Athlon processors with full-speed on-die L2 cache in the Coppermine and Thunderbird cores.

 

Other than that, Quake 3 shares its hunger for raw rendering power with its predecessor. For the Banshee chip this counts double, because it doesn't support multitexturing, which the game makes heavy use of.

 

The conclusion is straightforward: A small coppermine CPU is plenty enough for almost all Banshees. If you are after the last 1-2 FPS for your Gigabyte GA-630 or Diamond Monster Fusion, a 700 or 850 MHz Coppermine can help squeezing the last bit of performance out of these cards.


Conclusion


For a stock-clocked Banshee, as well as midly overclocked models, any 500 - 600 MHz Slot 1 CPU will suffice. If you are not after 3DMark highscores or play games based on the id Tech 3 engine, a 450 MHz Pentium II is also well suited for such a card. Most period-correct games won't gain more than 1 FPS with these basic models when switching to a faster processor, as can be seen in the benchmarks of UT99 or especially Quake 2. A good, cheap and readily available choice for this platform would be Pentium III Katmai CPUs with 450 or 500 MHz. Depending on what can be had cheaper, a Celeron 300A overclocked to 450 MHz is another excellent choice. Due to its on-die cache, it would probably perform just as good as a 500 MHz Katmai or even better.

 

For high-end Banshee models such as the Diamond Monster or the Gigabyte GA-630 in particular, a coppermine CPU with 550+ MHz is the better choice. These can squeeze the last bit of performance out of the cards and help boosting the framerate in more demanding games. The downside is that nowadays these CPUs (and boards supporting them) can be quite difficult and/or pricey to obtain, if you don't already have them in your collection.

 

That being said, I might do another comparison like this in the future, this time with the focus shifted on the comparison of the card across several different retro platforms. So stay tuned! 🙂

4 thoughts on “Benchmark – CPU Scaling 3dfx Banshee on Slot1”

  1. Danke für die ganze Arbeit.
    Interessant das der 600 mit dem Half-Speed Cache gegenüber dem 550 verliert…Der 300A@450 hat ja den 450 PII auch geschlagen…

    1. Gern 🙂
      Der Celeron 300A@450 hat ja zumindest gegenüber dem PII keinen Taktnachteil und dass kleinerer, schnellerer Cache die überlegene Technik sind, haben ja sowohl Intel als auch AMD mit Coppermine und Thunderbird verstanden und umgesetzt. Aber mit dem Vergleich Coppermine 550 vs Katmai 600 wird das Ausmaß nochmal deutlicher (zumindest bei “moderneren” Engines). Auf jeden Fall interessant zu sehen wie ich finde 🙂

Leave a Reply

Please use a Nickname for privacy.