Originally Posted by
101
I've done my two and the shader clock seems very unresponsive to voltage and is the main performance driver. I only went up to about 10% over volt tho, or about 1.3V. One of my cards could hit 1566 shader no problem, the other is stuck at 1458. The very next tick introduces instant artifacts, that can't be "dialed out" with more voltage as was possible on older cards.
Also, I noticed if you pushed the shaders really hard, the failure mechanism was very strange. Instead of just dropping vertices or pixels it would sometimes loose portions of the frame buffer altogether. When I say lost, I literally mean it's pointed to the wrong address in local memory. You could get it to show a flat 2D source texture as the entire screen, like if you browsed through the games assets. It was very interesting, particularly because the game (crysis/bioshock/shader heavy game) wouldn't be locked up. You could still escape out and get back to the desktop.
Based on that I think the 512bit architecture complexity is pushing the setup and hold time requirements for the bus to it's absolute limit. Particularly in keeping the shaders fed properly at increased clock rates. This would also explain how they are much more responsive to temperature than voltage and why 260s clock further, and looks to me like the typical limit to this current part. It certainly isn't starved for DC power.
Going to a smaller process will certainly help, as would reducing the bus width to relax the timing requirements. The only benefit Vcore gives you is to match the core clock to half the shader clock without issue, but the cards performance is largely driven by the shader clock.
Reply With Quote
Bookmarks