Second-hand CPU vs brand new GPU

[fivemack]

A bargainhardware.com dual E5-2650v1 machine costs £300, uses 220 watts, and does 7.2ms per iteration on four 44M exponents in parallel

A GTX1080 costs £665, uses about 150 watts, and does 2.4ms per iteration on a single 45M exponent.

So the machine designed five years ago for no-holds-barred double-precision wins quite handily even without the benefits of AVX2, and I should probably stop running cudalucas.

Running equivalent comparison for ECM now

[Gordon]

Originally Posted by fivemack

A bargainhardware.com dual E5-2650v1 machine costs £300, uses 220 watts, and does 7.2ms per iteration on four 44M exponents in parallel

A GTX1080 costs £665, uses about 150 watts, and does 2.4ms per iteration on a single 45M exponent.

So the machine designed five years ago for no-holds-barred double-precision wins quite handily even without the benefits of AVX2, and I should probably stop running cudalucas.

Running equivalent comparison for ECM now

Isn't the real question, why on earth would you run LL testing on a GPU? The 1080 is a beast at running TF.

[fivemack]

Originally Posted by Gordon

Isn't the real question, why on earth would you run LL testing on a GPU? The 1080 is a beast at running TF.

I get the impression there is more than enough GPU TF effort already in place. Mostly I got the GPU for factorisation, running ECM and polynomial selection, which it does pretty well.

[ATH]

GTX 1080 is one of those newer card with DP = 1/32th of SP performance, 257 GFLOPS vs 8228 GFLOPS, so it is not best for LL tests. There is really a need for a new consumer card with better DP performance.

[Mark Rose]

Apparently this year's Vega from AMD will have 1/16th DP.

That dual processor machine looks like a good deal for LL.

[Batalov]

Originally Posted by fivemack

.., I should probably stop running cudalucas.

Amen to that!

In contrast to cudalucas (which makes calls to some vanilla FFT and hits all the artificially NVIDIA-imposed bottlenecks), one should really want to run some DWT or NTT algorithm (like geneFer or Cyclo) to make GPUs really shine.

[Prime95]

Originally Posted by Batalov

In contrast to cudalucas (which makes calls to some vanilla FFT and hits all the artificially NVIDIA-imposed bottlenecks), one should really want to run some DWT or NTT algorithm.

CUDALucas does use a DWT algorithm.

I briefly explored an all-integer solution (not NTT though). My conclusion was I was unlikely to significantly beat the current CUDALucas timings. IIRC, it would be roughly +/- 20%. I think this was on a 6xx GPU.

[mackerel]

Originally Posted by fivemack

A bargainhardware.com dual E5-2650v1 machine costs £300, uses 220 watts, and does 7.2ms per iteration on four 44M exponents in parallel

They're taking offers at £250 nominal on ebay for the 32GB model. They seem to use 4GB 2R modules so you shouldn't run into ram bandwidth problems. When I got the 64GB model I offered a bit lower than their asking and they took it.

It would also be more fair to compare against previous generation used. Taking a 980Ti for example, it is approx 2/3 the rated boost SP FLOPS with a target cost under half a 1080, although TDP is higher at 250W.

If you really want DP, what about the R9 280X? It was possibly the last fast consumer card before they started to cripple DP. A quick look on ebay shows them under £150, and that gets you ball park of 1 DP TFLOP. Still 250W TDP though. If anyone can give me idiot proof instructions on how to bench it, I can do it on mine. I've bios mod lowered voltage so in practice it only takes around 200W now.

Originally Posted by ATH

GTX 1080 is one of those newer card with DP = 1/32th of SP performance, 257 GFLOPS vs 8228 GFLOPS, so it is not best for LL tests. There is really a need for a new consumer card with better DP performance.

Unfortunately probably not going to happen, unless you can find a compelling consumer DP requirement. If anything, the trend seems to be going the other way, with ever more FLOPS at lower precision.

[Batalov]

Right! (I must have been thinking about llrCUDA -- which simply calls FFTw)

NTT gave GeNeFer a new life (b ranges were extended, and now that it is implemented in OCL it is free of NVIDIA shackles).

[Mark Rose]

Originally Posted by mackerel

Unfortunately probably not going to happen, unless you can find a compelling consumer DP requirement. If anything, the trend seems to be going the other way, with ever more FLOPS at lower precision.

Rumours are Vega 20 will have 1/2 DP in 2018.

[pepi37]

Can llrCUDA can be rewritten as llrocl?