Host Interface Performance
With the introduction of PCI Express we've been testing the implementation on various graphics boards, both native chips and AGP chips bridged to PCI Express interfaces, to test the performance as PCI Express should offer considerable performance benefits. This is the first time that we've come across a PCI Express chip bridged back to AGP, so we'll again test to see what the bandwidth utilisation is on the boards we have already used in this article.
It should be noted that, theoretically, PCI Express for graphics is a superset of AGP's capabilities, so bridging a PCI Express back to AGP should be able to make the most of the bandwidth the host system can offer, although there may be some slight increase in latencies due to the inclusion of another chip in the data path.
Texbench (Downstream)
The first test we'll look at is "TexBench".This application allows you to test texture sizes with large number of textures. In this case we're testing up to 100 textures from sizes of 64x64 to 1024x1024, resulting in a maximum texture requirement of 400MB for storage, which is greater than the 128MB of frame buffer space available on any of the boards ontest here.
| Tex Bench Texture Size Download (MB/s) | 64x64 | 128x128 | 256x256 | 512x512 | 1024x1024 |
| 6600 GT AGP | 504.0 | 573.4 | 592.4 | 529.1 | 34.4 |
| 6600 GT PCIe | 504.0 | 672.0 | 696.4 | 646.4 | 135.5 |
| 5700 Ultra | 488.3 | 334.9 | 615.8 | 556.5 | 35.6 |
| 5800 Ultra | 488.3 | 664.9 | 639.4 | 551.9 | 36.5 |
| 6600 GT AGP % Faster Than: | 64x64 | 128x128 | 256x256 | 512x512 | 1024x1024 |
| 6600 GT PCIe | 0.0% | -14.7% | -14.9% | -18.1% | -74.6% |
| 5700 Ultra | 3.2% | 71.2% | -3.8% | -4.9% | -3.4% |
| 5800 Ultra | 3.2% | -13.8% | -7.3% | -4.1% | -5.9% |
The performance of the boards here are much as we'd expect with the native PCI Express 6600 GT having the greatest bus bandwidth utilisation when the texture sizes get bigger. Whilst the AGP 6600 GT initially has the same performance, the AGP version quickly drops back to similar utilisations of the other AGP boards, in face it has a slight deficit in most cases, which may be as a result of latencies.
Serious Magic (Upstream)
Another test that should highlight potential bus performance differences is Serious Magic'sTexture Download Benchmark. This application renders and image on the graphics board and then reads the image back to the host system, thus testing the upstream performance of the graphics board.
| Serious Magic Image Download (MB/s) | 640x480 | 800x600 | 1024x768 | 1280x1024 | 1600x1200 |
| 6600 GT AGP | 896.5 | 906.4 | 926.0 | 935.0 | 935.0 |
| 6600 GT PCIe | 836.0 | 893.0 | 910.0 | 923.0 | 929.0 |
| 5700 Ultra | 196.6 | 197.2 | 198.1 | 198.6 | 198.3 |
| 5800 Ultra | 196.6 | 197.3 | 198.3 | 198.9 | 199.0 |
| 6600 GT AGP % Faster Than: | 640x480 | 800x600 | 1024x768 | 1280x1024 | 1600x1200 |
| 6600 GT PCIe | 7.2% | 1.5% | 1.8% | 1.3% | 0.6% |
| 5700 Ultra | 356.0% | 359.6% | 367.4% | 370.8% | 371.5% |
| 5800 Ultra | 356.0% | 359.4% | 367.0% | 370.1% | 369.8% |
The results of the texture upload test are a little more inexplicable, in that the bridged AGP 6600 GT follows the performance not of the other AGP boards, but of the PCI Express 6600 GT - if fact, its even slightly higher!
Assuming these results are not an anomaly with this particular test one potential explanation could be how the HSI chip is initiating the transfers. AGP uploads can either operate at PCI Conventional rates or at the full 2.1GB/s - it may be that the test on the AGP boards is only operating at PCI Conventional rates, however the HSI bridge may be forcing the AGP to operate to full AGP rates on upstream data transfers.
