Thu Sep 11, 2008 12:25 pm
These WUs on RV770 almost double ppd. It's just a LOT more atoms, which doesn't scale as well on NV and scales loads better on the 4800.
It's all about the clocks of the SPs, the number of SPs, and the complexity of the SPs. ATI has kept clocks down, kept complexity medium, and jacked the number of SPs to eleven. NV kept the SP count low, the complexity medium/high, and keeps the SP clocks way high.
So with that, you can see why things have worked the way they have so far. NV was excelling on the WUs with small numbers of atoms, since the SPs were clocked way up. This is why ATI ppd were so low, the SPs are clocked around half of NV's. These new WUs have more atoms than the NV cards can calculate all at once, so they're having to loop back through while the ATI cards have enough SPs to get it all in one pass.
So before NV was doing everything in a single super fast pass and ATI was doing one slower pass.
Now NV is looping back and doing 2 or more super fast passes while ATI still does the same pass as before.
Pande has mentioned wanting to do bigger and bigger proteins with more atoms, so while this may not make for happy NV GPU folders, the ATI guys will be happy and more math will be done, just not as many points.
Please don't edit my signature for me. Thanks.