GPUGrid: Maxwell GTX750Ti

[RSS]

Hi Guys, Just a line to say we've got our first GTX750Ti card up and running. The performance out-of-the-box matches a GTX480. We'll have a public app for GPUGRID just as soon as CUDA 6.0 is publicly released, which will probably happen at GTC at the end of the month. Matt

More... (http://www.gpugrid.net/forum_thread.php?id=3651)

[zombie67]

Is "matching a GTX480" better or worse than expected, based on the specs?

[FourOh]

Based only on SP GFLOPS, they are a pretty close match (1345 vs 1306), but at 25% the power draw - the 750 ti seems promising to me. I have a spot for one when I can catch a deal.

Will the 128-bit memory architecture be a limiting factor?

[John P. Myers]

Will the 128-bit memory architecture be a limiting factor?
Nope. They've increased cache from 256kB to 2MB on the GM107 greatly reducing the number of memory fetches.

[Slicker]

128-bit memory is good, but if they could handle native 128 bit integers, it would quadruple the speed of the Collatz WUs compared to today! Collatz doesn't spend that much time transferring data to/from the host to the GPU. It may help when it reads the 4x32-bit integers (128 bits total) from the the lookup table though. It does that millions of times per kernel call. If it can read it in a single 128-bit chunk and map it to a uint4 structure, it will speed things up quite a bit. And, with a cache size that large, maybe the lookup table should be larger....

[FourOh]

I've been running the 750 Ti for a few days on GPUGrid - initially I was getting the new CUDA 6 long run work units, which I was able to complete with production of over 200,000/day. At 60W (max TDP) that's 3,333 credits/day/watt. Compare that to the GTX 580, which gets about 290k/day at 244W for 1189 credits/day/watt. So, two 750 Ti cards could do 40% more credits using half the power vs the GTX 580.

BUUUUUTT... after a couple days getting CUDA 6 long run work units, the project started sending CUDA 5.5 units, which all error out immediately. I'm only getting CUDA 6 short run units now, which means my production is only about 110k/day. Still better per watt than the 580, but not where I want to be credits-wise. I have some making up to do after that slog of an Einstein challenge!