What's going on at S3?

November 22, 2005 / by robbitop / page 3 of 4 / translated by the 3DCenter Translation Team

  deutsche Fassung dieses Artikels anzeigen

   Die aktuelle und zukünftige Modellreihe

Succeeding a short summary of past, current and announced graphicchip of S3:

With the Chrome S2x series, S3 Graphics majorly overhaul the Columbia design once more. The main focus of this architecture is clearly arithmetic power and not texturing performance. Unfortunately, Columbia's shader ALUs still cannot do co-issue, and that costs lots of performance. Furthermore S3 recognized that designing a control logic that could integrate multiple quad pipelines was too costly and time-consuming for Columbia and focused on only one quad pipeline, and making that highly efficient and scaling it to high clock speeds. Maybe for that reason GammaChrome S19 has never been released.

Despite some weaknesses, S3 Graphics deserves our respect for their second consequent expansion of an existing architecture. According to first benchmarks on the internet, the flagship of the Chrome S2x series, Chrome S27, is similar in performance to the Geforce 6600 DDR2 and the Radeon X1300 Pro - a really competitive result.

   Performance examination DeltaChrome vs. GammaChrome

Despite our continued efforts, we still do not have a GammaChrome S18 review sample, so our analyses had to be based on results from Beyond3D and our own tests of DeltaChrome S8. Only synthetic tests have been taken into account, to isolate graphics performance from any system influences. However, without having access to a review sample, we must state that these results are to be taken with a pinch of salt.

One thing up front: Despite having only half the pipelines, the GammaChrome S18 is faster than the DeltaChrome S8. How can that be? Are only the aforementioned chip-internal improvements responsible? To figure this out, we have to again take a closer look at some things. Therefore, we examine GammaChrome S18 Nitro: It operates at 450 MHz core and memory clocks, as opposed to DeltaChrome S8's 300/300 MHz.

When you calculate raw fillrate of both graphics chips, DeltaChrome S8 has an approximate one third advantage. However, DeltaChrome can fully use this fillrate only in very rare cases. Firstly because of the aforementioned losses to the simplistic quad pipeline control logic, secondly because of severe lack of bandwidth.

GammaChrome on the other hand does not suffer from these problems. It has got about 50% more bandwidth, and can use it more efficiently thanks to its two-channel partitioned memory controller. Effective fillrate of GammaChrome S18 should be equal or even higher than those of DeltaChrome S8 depending on the situation (size of triangles, number of texture layers).

Today's IMRs can skip many occluded objects without actaully rendering them, which increases efficiency. Details can be found in a previous article. This occlusion culling has a large impact on real game performance. "GL_EXT_REME" (a program by Humus) is a synthetic benchmark that measures Z rejection performance.

HSR unit and ROPs are responsible for Z rejection. In "traditional" pipeline architectures like the DeltaChrome, it scales with number of pixel pipelines and clock. Remarkably, the GammaChrome has got an enormously increased Z rejection performance despite its lower raw fillrate. We attribute this to a now fully working hierarchical Z buffer, which can discard lots of pixels at once. This was not yet the case with DeltaChrome.

For calculating realistic shadows, stencil shadows are widely used today. Games like Riddick and Doom 3 are examples. Excessive use of stencil shadows certainly requires high stencil performance. This significantly depends on the abilities and number of ROPs in a GPU. In a previous article we observed DeltaChrome's low stencil performance, despite its 8 ROPs, by using Fable Mark.

To apply stencil shadows effectively and reduce overdraw, many renderers use a "Z first" pass, during which the Z buffer is prerendered. There are now two render passes instead of one, but a huge part of overdraw is eliminated - so the main render pass gets less complex. To ensure a fast Z pass, a GPU must have a high Z performance. Both of this is used excessively in game test 2, "Battle of Proxycon", in 3DMark2003. Consequently this can act as a good indicator for Z and stencil performance of a graphics chip.

GammaChrome seems to feature drastically improved ROPs, concerning z and stencil performance. Normalized for clock, a GammaChrome pipeline is about 3 times as fast at handling z and stencil in Game Test 2 of 3DMark03 as a DeltaChrome pipeline. This sounds like a lot, but this only brings a previous critically weak spot up to normal standards.

The increased core clock can almost balance out the raw performance. Add to that the significantly higher efficiency of GammaChrome over DeltaChrome. Much less of its potential goes wasted because there is no second quad pipeline to battle over a triangle with. Furthermore the bandwidth-to-fillrate ratio has risen over DeltaChrome S8, so that GammaChrome S18 can translate an appreciably higer portion of its raw fillrate into framebuffer writes. Real-world fillrates of GammaChrome should be higher than those of DeltaChrome. Pixel shader and vertex shader performance have also risen. And the aforementioned ugly bugs got fixed, so that lots of situations are going to work more smoothly.