Nvidia Kepler

[KyleAskine]

You will get the reports from Jerry once he gets it running I am sure. But it really sounds like this is a gamer card, and not a compute card.

[flashjh]

Originally Posted by pinhodecarlos

Just by looking at cards TDP I don't understand your worries. In my point of view, energy efficiency one, making at least the same with less energy is good.

What we're concerned about isn't TDP... to be honest, in the mfakt* and CUDALucas, et al. business, if you're concerned about power consumption, you're doing it for the wrong reasons . Really, the tripple CUDA cores looked really exciting, but now with some real reviews and other info discussed here it looks like the 680 may not be much better than a 580 for this type of work.

Originally Posted by bcp19

That said, I'm eagerly awaiting flash's report on the performance of the 680 in (hopefully) both mfaktc and CUDALucas.

Originally Posted by KyleAskine

You will get the reports from Jerry once he gets it running I am sure. But it really sounds like this is a gamer card, and not a compute card.

I'm excited to get it, shipped today. I'm also skeptical because all the initial info said it would be twice as fast as a 580; now, I'm not so sure. Either way, I'll have the system ready for mfaktc and CUDALucas testing and I'll post ASAP once I get it in. (UPS still doesn't have a delivery date)

Edit: One has to wonder, if this card doesn't live up to the actual 680 name, what is nVidia thinking? I know I can't afford the Tesla and Quadro cards, and based on their performance, I wouldn't really want one anyway. If Kyle is on the right track, I can't imagine what we're going to do. On the other hand, nVidia and AMD can both see where the technology is going... CUDA and OpenCL have enough of a backing that they have to know why some of us are buying the cards. Seems to me they're only hurting themselves if they don't make products that live up to the hype.

[KyleAskine]

Originally Posted by flashjh

What we're concerned about isn't TDP... to be honest, in the mfakt* and CUDALucas, et al. business, if you're concerned about power consumption, you're doing it for the wrong reasons . Really, the tripple CUDA cores looked really exciting, but now with some real reviews and other info discussed here it looks like the 680 may not be much better than a 580 for this type of work.

Agree on all accounts. If you want more shaders you can always go to AMD, they have boatloads of shaders, but it doesn't really seem to help them run mfakto.

Originally Posted by flashjh

Edit: One has to wonder, if this card doesn't live up to the actual 680 name, what is nVidia thinking? I know I can't afford the Tesla and Quadro cards, and based on their performance, I wouldn't really want one anyway. If Kyle is on the right track, I can't imagine what we're going to do. On the other hand, nVidia and AMD can both see where the technology is going... CUDA and OpenCL have enough of a backing that they have to know why some of us are buying the cards. Seems to me they're only hurting themselves if they don't make products that live up to the hype.

If you read the Tom's Hardware feature today on the 680, they more or less come out and say what I am saying (don't think I am smart enough to divine nVidia's business plans on my own). It really sounds like the 680 will be very underwhelming from a compute standpoint, and that this is intentional. I hope I am wrong and we get a 50% improvement, but we will know next week.

AMD really does seem to be making a strong move towards the compute space, which makes this decision all the more curious.

[nucleon]

At my level, I'm definitely concerned on the work/power consumption :)

I had to shift one of the PCs onto a different power circuit.

Rough figures, if I replace all my 580s/560Tis with 680s, I'd be looking at 300W power saving - enough to run another one & partial CPU :)


-- Craig

[ET_]

Originally Posted by nucleon

At my level, I'm definitely concerned on the work/power consumption :)

I had to shift one of the PCs onto a different power circuit.

Rough figures, if I replace all my 580s/560Tis with 680s, I'd be looking at 300W power saving - enough to run another one & partial CPU :)


-- Craig

I'm (morer than) a bit concerned about the GTX 680. Apart from gaming and video encoding, it seems that the card has been "tailored" for gaming platforms with a price near the high-end devices.

I'm seriously considering a GTX 580 as soon as the price rebates... but first, I'm waiting for some tests.

Luigi

[Karl M Johnson]

Originally Posted by ET_

I'm (morer than) a bit concerned about the GTX 680. Apart from gaming and video encoding, it seems that the card has been "tailored" for gaming platforms with a price near the high-end devices.

I'm seriously considering a GTX 580 as soon as the price rebates... but first, I'm waiting for some tests.

Luigi

Yep, looks like it.
ATM, GTX 680 behaves like a sm_21 GPU (though it should be sm_31 ?) - it's great at gaming, but it's worse than sm_20 GPUs in FP, DP FP and INT ops.
However, mostly FP and DP FP tests were done in reviews in compute sections, so yeah, we need our own tests to make final judgement.
Would be a shame if it turns out to be a pure gaming GPU (which means NV turns greedy and encourages compute folks to use Teslas).

[bcp19]

Originally Posted by KyleAskine

Agree on all accounts. If you want more shaders you can always go to AMD, they have boatloads of shaders, but it doesn't really seem to help them run mfakto.

I think this is probably more due to the CUDA vs OpenCL differences. CUDA is better able to take advantage of resources than OpenCL. If you look at the ratings of cards on JamesH's site, the 6990 has 5.1 GFlops capability vs 1.6 on the 580, yet the 580 outperforms it by over 10% on TF.

[Karl M Johnson]

Ofc NV's native programming interface is better than OpenCL.
However, even if CUDA apps on gtx 680 will be 10-20% faster, it's still a crappy result (in terms of performance per shader per Ghz).

I like to measure peak theoretical performance in int ops rather than fp, since sometimes it's possible to "cheat" by multiplying the performance by 2 or 3, while peak theoretical performance with integers is always shaders * SD clock.

[Karl M Johnson]

If the chinese will keep their promise, I will get that gpu tomorrow!
I AM SO ANXIOUS!
The mystery of it's performance is driving me nuts.

[firejuggler]

don't except anything, this way you will always be pleasantly surprised.
If it hold its promise -as in 3 time the compute power of a 580- you won't be able to saturate it. you will need at least 2 of the new xeon.

[lalera]

hi,
mfaktc v0.18 does not run with gtx680

mfaktc v0.18 (64bit built)

Compiletime options
THREADS_PER_BLOCK 256
SIEVE_SIZE_LIMIT 32kiB
SIEVE_SIZE 193154bits
SIEVE_SPLIT 250
MORE_CLASSES enabled

Runtime options
SievePrimes 25000
SievePrimesAdjust 1
NumStreams 3
CPUStreams 3
GridSize 3
WorkFile worktodo.txt
Checkpoints enabled
CheckpointDelay 30s
Stages enabled
StopAfterFactor bitlevel
PrintMode full
AllowSleep no

CUDA version info
binary compiled for CUDA 4.10
CUDA runtime version 4.10
CUDA driver version 4.20

CUDA device info
name GeForce GTX 680
compute capability 3.0
maximum threads per block 1024
number of mutliprocessors 8 (unknown number of shader cores)
clock rate 705MHz

Automatic parameters
threads per grid 1048576

running a simple selftest...
ERROR: cudaGetLastError() returned 8: invalid device function