Well judge for yourselves by analyzing the scores that we got by overclocking our board. Clearly there's not too much headroom available for overclocking and you may or may not be able to do so. We'd suggest increasing in increments of 5MHz at a time. Our DDR-SGRAM was able to go to a frequency as high as 135MHz for the graphics clock. Any higher and there were a few discrepancies with some of the benchmarks. In Quake 3's High Quality test, we managed to eke out an extra 4 fps at 135/311MHz, which might not seem like a lot but at least the opportunity is there for those of you that want to grasp it.
One final word of warning, with 23million transistors the chip gets a bit on the hot side, ok? So don't expect too much. You could get lucky with your memory type though, but if you're thinking of doing something 'clever' with the DDR-SGRAM, um… think again. At a default 301MHz we pushed it up past 311MHz and it didn't seem to respond to well. It got a little sick and showed a few spots and the like.
Ok so what can we tell you… obviously September 27th has been and gone and without Intel's Camino (i820). The RDRAM issue has pushed Intel's proposed 820 chipset with its AGP 4X to a date that isn't yet set in stone (um… to use id Software's quote 'when it's done'). In our original preview, we mentioned that NVIDIA's usage of AGP 4X and in particular 'Fast Writes' should add some much needed bandwidth for shifting those large textures (as opposed to 3dfx proprietary DME transfer). When scouring through their white papers last time around on AGP 4X, we noticed that they had designed their own AGP Fast Writes north bridge (per AGP 2.0 spec) and used it as a test vehicle for the GeForce 256 GPU. Don't worry though 'Fast Writes' still works with AGP 2X of course. The 'Fast Write' feature is an AGP 2.0 specification. It allows the CPU to send data directly to the graphics bus without having to go through system memory (removing a bottleneck) and it also frees up the system memory bus allowing it to perform other functions. Fast Writes improves performance on certain calculations from the CPU to the GPU, which include these four main areas:
- 2D operations
- Operations involving writing data to the frame buffer or sending data to the graphics chip
- Loading textures in D3D into local memory
- Writing push buffers to graphics local memory (should be where the greatest benefit of the benefit will be had).
Still with Intel's AGP 4X perhaps VIA's own AGP4X implementation will spice things up for 'Fast Writes'. We'll keep you posted. In the meantime it's back to 'theory'. Taking up NVIDIA's claims that it takes 90bytes of bandwidth per triangle and again assuming that 10million is the ceiling, that would soak up some 900Mbytes
