What's going on at S3?
November 22, 2005 / by robbitop / page 4 of 4 / translated by the 3DCenter Translation Team
In the following round up we are going to compare both architectures' efficiency in terms of pixel shading, in order to see how S3 has progressed in this regard. To this end we need to normalize the results to factor out the difference in pipeline count. DeltaChrome sports 8 pipelines, while GammaChrome has only 4, but has a higher clock frequency: 450 vs. 300 MHz.
So the following is a comparison of performance per pipeline and clock cycle. We used Shadermark. This synthetic test has only one bottleneck: Shader performance. The rest of the system is irrelevant. Therefore, in this special case, we could pit Beyond3D's results against our own.
| Shadermark FP16 | DeltaChrome S8 Pro | GammaChrome S18 Nitro | Performance gain per pipeline |
| Shader 2 | 161 fps | 138 fps | 13.6% |
| Shader 3 | 102 fps | 87 fps | 12.7% |
| Shader 4 | 95 fps | 83 fps | 15.7% |
| Shader 5 | 77 fps | 67 fps | 15.5% |
| Shader 6 | 111 fps | 97 fps | 16.2% |
| Shader 7 | 86 fps | 72 fps | 11.6% |
| Shader 8 | 51 fps | 51 fps | 33.3% |
| Shader 9 | 233 fps | 216 fps | 23.1% |
| Shader 10 | 177 fps | 167 fps | 25.4% |
| Shader 11 | 99 fps | 87 fps | 17.1% |
| Shader 12 | 41 fps | 43 fps | 39.0% |
| Shader 13 | 49 fps | 53 fps | 42.8% |
| Shader 14 | 32 fps | 33 fps | 37.5% |
| Shader 15 | 32 fps | 28 fps | 15.6% |
| Shader 16 | 55 fps | 52 fps | 25.4% |
| Shader 17 | 52 fps | 49 fps | 25.0% |
| Shader 18 | 6 fps | 6 fps | 33.3% |
| Shader 19 | 17 fps | 18 fps | 41.1% |
Shadermark's default FP24 mode reveals even bigger gains:
| Shadermark FP24 | DeltaChrome S8 Pro | GammaChrome S18 Nitro | Performance gain per pipeline |
| Shader 2 | 145 fps | 128 fps | 26.2% |
| Shader 3 | 98 fps | 87 fps | 17.3% |
| Shader 4 | 89 fps | 83 fps | 23.5% |
| Shader 5 | 73 fps | 67 fps | 21.9% |
| Shader 6 | 106 fps | 97 fps | 21.6% |
| Shader 7 | 81 fps | 72 fps | 18.5% |
| Shader 8 | 24 fps | 51 fps | 183.3% |
| Shader 9 | 194 fps | 216 fps | 47.9% |
| Shader 10 | 159 fps | 167 fps | 39.6% |
| Shader 11 | 93 fps | 87 fps | 24.7% |
| Shader 12 | 40 fps | 43 fps | 42.5% |
| Shader 13 | 26 fps | 53 fps | 169.2% |
| Shader 14 | 26 fps | 33 fps | 69.2% |
| Shader 15 | 18 fps | 28 fps | 105.5% |
| Shader 16 | 31 fps | 52 fps | 122.5% |
| Shader 17 | 28 fps | 49 fps | 132.1% |
| Shader 18 | 4 fps | 6 fps | 100.0% |
| Shader 19 | 10 fps | 18 fps | 140.0% |
As you can see, for a refresh, the increase in efficiency is respectable, especially in FP24 mode. A revved up GammaChrome with a second quad-pipe and the appropriate clock frequencies (as planned with Chrome S25/S27) could breathe new life into the original Columbia design. Chrome S2x will then see further improvements in performance per clock and pipeline. As we do not have a test sample, we are unable to verify this claim for the time being, but will certainly do so as soon as possible.
Driver progress
Since DeltaChrome has been introduced, its drivers have made a fair amount of progress. In October 2003 the number of playable games was quite limited. Performance was not great either. The OpenGL driver seemed to be a big construction ground, with most extensions missing. Following the subsequent releases of Far Cry, Half-Life 2 and Doom 3 both the Direct3D and OpenGL parts of the driver seem to have gotten bigger revamps. Performance and compatibility have greatly improved. Nearly each new driver release is accompanied by a number of game fixes and, from time to time, there are also performance enhancements.
CPU usage however has been a problem since the beginning, the driver consumed lots of CPU cycles. Hence nearly all titles were CPU bound - in nearly all resolutions. This has been somewhat improved meanwhile, but is still apparent in comparison to the big competition. S3 has confirmed that CPU usage indeed is an issue, but is caused by a bug in DeltaChrome's design, and that bug, according to S3, is fixed with GammaChrome.
Yours truly, who has been using a DeltaChrome for one year and a half, believes that the drivers have improved significantly. There are nearly no concessions left in the daily usage of the card. Driver development is painful in the the early stages of a new architecture. ATI's R300 and nVidia's NV30/40 show that even big players are not immune against this. S3 with their relatively small driver development team remain at a disadvantage versus ATI and nVidia. However, as Chrome 2x belongs to the same Columbia architectural family as the current offerings, DeltaChrome and GammaChrome, there is hope that the new chip will have a relatively mature driver package to go with at launch.
To show the progress in driver development we have compiled a quick comparison with the following test system:
- Intel Pentium 4 3000 @ 3600 MHz (FSB240)
- 2 x 512 MB DDR333 @ DDR384, 2.5-3-3-8
- S3 DeltaChrome S8 Pro
- Drivers: 15.11.06j, 15.13.04f1, 15.13.04h, 15.15.08c & 15.16.07
The results in the table are in chronological order of the driver versions:
| 1024x768 | 1024x768 8xAF | 1280x1024 | 1280x1024 8xAF | |
| Call of Duty Demo | 66/63/68/48/83 | 61/63/59/39/71 | 54/53/54/38/61 | 44/46/45/33/46 |
| Colin Rally 5 Demo | 23/24/23/31/30 | 19/19/20/24/25 | 16/16/16/20/20 | 14/14/14/16/17 |
| Doom 3 | 11/10/10/10/10 | 10/10/10/10/10 | 7/7/7/7/7 | 7/6/6/6/6 |
| Far Cry | 13/14/16/14/16 | 5/14/14/14/14 | 8/11/11/11/11 | 3/9/9/10/10 |
| Half-Life 2 | 24/20/20/21/23 | 22/19/19/19/21 | 16/14/14/14/16 | 15/13/12/13/14 |
| Max Payne 2 Demo | 83/84/83/92/80 | 68/67/68/79/68 | 55/56/55/62/56 | 47/47/48/54/48 |
| Painkiller SP2-Demo | 53/60/60/56/58 | 36/36/36/37/37 | 43/44/43/44/44 | 26/26/26/26/26 |
| Riddick Demo | -/-/-/9/9 | -/-/-/9/9 | -/-/-/6/6 | -/-/-/6/6 |
| Serious Sam Demo | 86/109/109/112/112 | 60/72/72/73/72 | 57/68/68/69/69 | 40/45/46/46/46 |
| Trackmania Sunrise | -/52/52/-/49 | -/44/44/-/44 | -/43/43/-/42 | -/36/37/-/36 |
| UT2004 Demo | 60/53/52/55/64 | 54/52/52/49/54 | 56/51/51/53/58 | 41/41/41/42/42 |
As can be seen from the above, S3's drivers have been considerably improved, in terms of both compatibility and speed.
Die Zukunft bei S3
With the introduction of Chrome 2x at the beginning of next year, S3 might be in a position to increase their market share. Although the new chip is still based on the GammaChrome design, S3 will finally offer updated chips with two quad-pipelines. On top of that the clock speed is supposed to be considerably higher thanks to Fujitsu's new 90nm process.
The Chrome S25 is intended as mainly an OEM product, so there should be a number of versions with a 64 bit memory interface only. Chrome S27, however, will be pitted against the GeForce 6600 as well as the Radeon X700 and X1300, bearing a price tag of around 100 USD/EUR. S27's planned clock speeds of 700/700 MHz promise to make it a very competitive mainstream offering.
Like ATI and nVidia, S3 plans to introduce a multiple-board technology as well. Chrome S27 will feature the SLI/CrossFire alternative "MultiChrome", which will allow the parallel usage of two graphics boards. Contrary to its competitors, S3 is willing to cooperate with all chipset vendors, and plans to make "MultiChrome" available on all mainboard chipsets. This might mean that, best case, MultiChrome will work with all mainboards with two PEG slots. We believe that MultiChrome will be based on a simple AFR method only.
Coupling two Chrome S27 cards via MultiChrome will deliver 5,6 Gigapixels of trilinear fill rate, plus a memory bandwidth of 44,8 GiB/s, which is fairly high for a mainstream offering. This might be enough to enable 5x super sampling anti-aliasing in combination with the anisotropic filter (of course depending on the fact that the two S27 chips will deliver playable frame rates in this mode).
To cut a long story short: S3 has good chances to achieve favourable results in the graphic chips markets as of next year. This however may be put at risk if - like already happened in the past - S3 will be unable to deliver according to their schedule to have Chrome 2X available at the beginning of next year. Should there be new delays, ATI and nVidia will in the meantime launch new products again, and S3 will simply no longer be able to sell the Chrome 2X chip in quantities.
In the long run there will be the Direct3D10 project "Destination films" which has been in the project phase since a long time now. Due to the delays caused by Microsofts rescheduled Windows Vista launch, this will not be relevant before H2/2006, anyway not before ATI and nVidia will have launched there own Direct3D10 parts. Up to date there are no technical specs to "Destination Films" available, with the only known facts being Direct3D10 and Shader 4.0.
At the end of this article I want to thank a few persons who helped me a lot with this. This translation of it is done by Hakkerstiwwel, zeckensack, Banshee18. For help in the driver test my thank goes to misterh. Last but not least I thank aths and Xmas for their clues and a few ideas which improved this article a lot. Thank you very much guys!
|
|
