Does anyone know how much is the double precision peak for GTX 280 and new Tesla 10 cards.
For GTX 280, 30 double precision units are mentioned which can do MAD every cycle.
This gives 3021.3 = 75 Gflops approx. Is this correct?
Tesla 10 also has the same architecture and same peak DP?
A German news site is mentioning a peak performance of 90 GFlops (Nvidia kratzt mit neuen Grafik- und Compute-Prozessoren an der TFLOPS-Schallmauer | heise online)
So your 75 GFlops are reasonable.
I’m very disappointed in the DP performance. It’s only 1/8 of the SP perfomance.
Yeah, more power is always better, but I’m impressed/surprised that double precision even made it into the consumer/gamer cards. That is an interesting design choice, which says (to me anyway) that NVIDIA expects things like PhysX to become a big deal this next generation.
(Or maybe it says they only wanted to fab one kind of chip to keep Tesla prices down…)
Since the dawn of time (at least in old-school GPGPU), people complained that GPGPU was pointless due to the lack of double precision. Now double precision surfaces on GPUs, and people complain that it’s too slow…
We have to rethink our algorithms for massively parallel execution anyway (or does anyone seriously believe that a 16-core Opteron/Xeon/ItaniumX 1-2 years from now will be as fast on a single core as top-of-the-line singlecore processors now?). If we invest time to rethink our algorithms, we might just spend a few more days thinking about the numerics, about some error distributions and how they build up in floating point math, and design our algorithms to only use the kind of precision necessary for each step. At least in the linear algebra domain (blas 1/2/3 & lapack), this is a solved problem. Such mixed precision approaches are even beneficial on today’s CPUs…
Anyway, I don’t believe anyone “who knows” and this includes all NVIDIA employees will publicly speculate on next-gen Tesla performance here. I am surprised that the Heise folks provide numbers, so far, they had a reputation of being trustworthy.
Well, since the reviewer NDAs on the GTX 280/260 lifted today, all of the reviews you are reading on the web now are hopefully informed either by NVIDIA literature or actual hardware.
(My apologies to all for this semi-private message. I have no other way to send this.)
Dominik - A forum/browser bug (probably relating to the non 7-bit ASCII character in your user name) is preventing me from replying to your private message, saying your user does not exist. If you PM me your email address I can reply directly.
(I now return you to your normal public forum)
Sorry for my ancestors to put a german umlaut in their family name ;) My account on these forums is slightly flawed anyway, some copying of user data went rogue when I got promoted from private to public forums. I think now I know what went wrong…
I was just told that any information on the Geforce and Tesla doubleprec GPUs is no longer disclosed as of today, so my point is no longer valid anyway. In particular, I continue to trust the folks over at Heise.
Well it IS a little slow given that Radeon 3870, a card that costs 1/3rd as much does about 100 Gflops DP.