TF benchmarks, mprime vs GPU

[rudi_m]

Hi,

since mprime 29.1 I'm really excited about comparing x86 vs GPU trial factoring benchmarks again. Also I don't have a AVX-512 machine and would like to see what would be the benefit of Xeon over consumer i5/7

Here I start with a real job benchmark, using a plain i7-6700 CPU. The table below shows the times needed to progress 1.27% of factoring M701000023 from 78 to 79 bit.

The columns show
1. used CPU cores
2. with/without hyper threading
3. speed factor regarding single thread
4. absolute time in seconds for 1.27% progress (OutputIterations=500000)
5. hyper threading benefit in percent

1 cores -ht x1.00 4543.147
1 cores +ht x1.22 3714.403 (+22.3%)
2 cores -ht x1.99 2285.129
2 cores +ht x2.43 1869.446 (+22.2%)
3 cores -ht x2.94 1545.755
3 cores +ht x3.59 1264.282 (+22.3%)
4 cores -ht x3.89 1168.383
4 cores +ht x4.65  977.141 (+19.6%)

As you see the multi-threading benefit is almost perfectly linear. (3 and 4 threads may look a bit worse because I had many other processes running on the server and also because of "less turbo GHz")

So now I'm curious how would AVX-512 or GPU perform for a similar job.

[TheJudger]

Stock GTX 1060 is around 4h 30m for M701000023 from 2⁷⁸ to 2⁷⁹.
Your 977s for 1.27% translates to ~21h 20m.
So don't waste your CPU time with TF and do some LL tests, we (GIMPS) are doing way to much TF already.

Oliver

[rudi_m]

Originally Posted by TheJudger

Stock GTX 1060 is around 4h 30m for M701000023 from 2⁷⁸ to 2⁷⁹.
Your 977s for 1.27% translates to ~21h 20m.

Ok, this is still factor 4.7 faster than CPU (without AVX-512) but estimated only factor ~2.3 better regarding power consumption.

Originally Posted by TheJudger

So don't waste your CPU time with TF and do some LL tests, we (GIMPS) are doing way to much TF already.

Yes Sir ;) Only have to finish my project to write a "column of one's" into the "P-1 Available" field on the "work distribution map" https://www.mersenne.org/primenet/

It's all about fun ;)

[rudi_m]

BTW in case one wants to help to safe my CPU power for LL you may pick up these for jobs for me :

Factor=602000051,78,79
Factor=603000011,78,79
Factor=604000003,78,79
Factor=605000003,78,79
Factor=606000001,78,79
Factor=607000003,78,79
Factor=608000017,78,79
Factor=609000137,78,79
Factor=611000003,78,79
Factor=612000029,78,79
Factor=613000009,78,79
Factor=614000021,78,79
Factor=615000019,78,79
Factor=616000127,78,79
Factor=617000107,78,79
Factor=618000017,78,79
Factor=619000009,78,79
Factor=621000007,78,79
Factor=622000039,78,79
Factor=623000033,78,79
Factor=625000069,78,79
Factor=626000047,78,79
Factor=627000053,78,79
Factor=628000057,78,79
Factor=629000011,78,79
Factor=631000001,78,79
Factor=632000011,78,79
Factor=633000013,78,79
Factor=634000019,78,79
Factor=635000089,78,79
Factor=636000031,78,79
Factor=637000009,78,79
Factor=638000003,78,79
Factor=639000083,78,79
Factor=641000117,78,79
Factor=642000031,78,79
Factor=643000031,78,79
Factor=644000011,78,79
Factor=645000007,78,79
Factor=646000049,78,79
Factor=647000017,78,79
Factor=648000083,78,79
Factor=649000073,78,79
Factor=651000047,78,79
Factor=652000003,78,79
Factor=653000003,78,79
Factor=654000059,78,79
Factor=655000097,78,79
Factor=656000003,78,79
Factor=657000121,78,79
Factor=658000139,78,79
Factor=659000137,78,79
Factor=661000117,78,79
Factor=662000011,78,79
Factor=663000031,78,79
Factor=665000029,78,79
Factor=666000019,78,79
Factor=667000001,78,79
Factor=668000131,78,79
Factor=669000047,78,79
Factor=671000017,78,79
Factor=672000013,78,79
Factor=673000037,78,79
Factor=674000023,78,79
Factor=675000037,78,79
Factor=676000007,78,79
Factor=677000201,78,79
Factor=678000031,78,79
Factor=679000019,78,79
Factor=681000013,78,79
Factor=682000063,78,79
Factor=683000053,78,79
Factor=684000067,78,79
Factor=685000003,78,79
Factor=686000009,78,79
Factor=687000049,78,79
Factor=688000009,78,79
Factor=689000027,78,79
Factor=691000099,78,79
Factor=692000017,78,79
Factor=693000109,78,79
Factor=694000171,78,79
Factor=695000003,78,79
Factor=696000013,78,79
Factor=697000013,78,79
Factor=699000017,78,79

[storm5510]

Originally Posted by TheJudger

... we (GIMPS) are doing way to much TF already.

Oliver

I've been using mfaktc since this past November. I can run about 173 GHz days per day on this hardware (GTX-750 Ti). This is in the 146 million to 152 million range.

I did a double-check test with an exponent in the 44 million range. CuLu estimated completion in 6 days and 19 hours. Prime95 indicated 6 days and 9 hours. I was a bit baffled that Prime95 could do this faster. Note: I did not make any changed to the CuLu configuration file.

Should I being doing other work or continue with mfaktc?

Thanks!

[VictordeHolland]

Originally Posted by storm5510

I've been using mfaktc since this past November. I can run about 173 GHz days per day on this hardware (GTX-750 Ti). This is in the 146 million to 152 million range.

I did a double-check test with an exponent in the 44 million range. CuLu estimated completion in 6 days and 19 hours. Prime95 indicated 6 days and 9 hours. I was a bit baffled that Prime95 could do this faster. Note: I did not make any changed to the CuLu configuration file.

Should I being doing other work or continue with mfaktc?

Thanks!

GPUs excel at TF, but for LL they're about the same as a quad-core CPU.

[ATH]

Originally Posted by storm5510

I did a double-check test with an exponent in the 44 million range. CuLu estimated completion in 6 days and 19 hours. Prime95 indicated 6 days and 9 hours. I was a bit baffled that Prime95 could do this faster. Note: I did not make any changed to the CuLu configuration file.

The problem is LL test requires double precision while TF only uses single precision. If you look at your GTX 750 Ti here:
https://en.wikipedia.org/wiki/GeForc...eries#Products

You can see it does 1306 GFLOPS single precision but only 40.8 GFLOPS (1/32 * 1306) in double precision, that is why LL test is so slow compared to TF.

Almost all "consumer" graphic cards have DP = 1/32 * SP or DP = 1/24 * SP except the original Titan, Titan Black and Titan Z from 2013/2014, they have DP = 1/3 * SP. You can see them at the bottom of that same list.

[Mark Rose]

The GTX 750 Ti isn't a particularly powerful GPU. A 780 or 580 is about 4 times faster for LL. A 1080 is about 7 times faster.

The project is getting more than enough TF work to keep ahead of LL. I understand we're only a little ahead of the P-1 work. We're probably two decades away from the 146M to 152M range.

The GTX 750 Ti may not be the world's fastest GPU, but put it this way: you can get double your LL throughput by using it for LL, so why not.

[storm5510]

Originally Posted by Mark Rose

The GTX 750 Ti isn't a particularly powerful GPU. A 780 or 580 is about 4 times faster for LL. A 1080 is about 7 times faster.

The project is getting more than enough TF work to keep ahead of LL. I understand we're only a little ahead of the P-1 work. We're probably two decades away from the 146M to 152M range.

The GTX 750 Ti may not be the world's fastest GPU, but put it this way: you can get double your LL throughput by using it for LL, so why not.

The 580 and 780 are way outside my cost range.

The CPU in this machine is an i5-3570 quad @ 3.4 GHz. Under a load, it will go up to 3.7. I've done pairs of P-1 tests with Prime95 in about 20 hours using two worker windows. The other two cores assist. This system can use an i7. So far, I haven't seen any reason to change it.

Now, I'll ask a question that I've been wondering about: What is the connection between TF and P-1, ECM, and so on?

[Mark Rose]

Originally Posted by storm5510

The 580 and 780 are way outside my cost range.

Shouldn't be. I've bought a used GTX 580 as cheap as $50 here in Canada. That being said, it consumes far more power than a GTX 750 Ti. Gaming performance between the two will be about the same.

The CPU in this machine is an i5-3570 quad @ 3.4 GHz. Under a load, it will go up to 3.7. I've done pairs of P-1 tests with Prime95 in about 20 hours using two worker windows. The other two cores assist. This system can use an i7. So far, I haven't seen any reason to change it.

There's no point getting an i7 for Prime95.

Now, I'll ask a question that I've been wondering about: What is the connection between TF and P-1, ECM, and so on?

GP2 wrote an excellent post answering that.

[Gordon]

Originally Posted by Mark Rose

Shouldn't be. I've bought a used GTX 580 as cheap as $50 here in Canada. That being said, it consumes far more power than a GTX 750 Ti. Gaming performance between the two will be about the same.

Remembering that a 580 on full power sounds like a jet engine taking off and the power consumption....