Prime95 version 28.6 / 28.7 (28.7 now available!)

[aurashift]

Can you add 100M exponent TF as a job type, please? Watching a factor come in on an 11-core worker makes me very sad.

[Prime95]

Originally Posted by aurashift

Can you add 100M exponent TF as a job type, please? Watching a factor come in on an 11-core worker makes me very sad.

Watching prime95 do TF makes me very sad. We need to have GPU72 build up a reserve of 100M exponents that are already TFed.

[chalsall]

Originally Posted by Prime95

We need to have GPU72 build up a reserve of 100M exponents that are already TFed.

For clarity, as you asking me to bring back into GPU72 the ~332.2M range for additional TF'ing, keeping the highly TF'ed candidates with Primenet for LL assignment? Or do you really mean 100M?

Easy to do either (or both) -- about 30 seconds of human work for the former; about 5 minutes for the latter. Just tell me what you'd like done.

[manfred4]

There already is such a Job, but since that link is not available from the site itself, not many people use it.
Since you said you were doing lots of work for LL in the 100M range now, and since i finished my last big Job (bringing 20M-22M from 66 to 67) I already am doing the recommended work in the 100M range.

GPU72 once had the 100M range TF option, but it disappeared some time ago. Maybe it should be added back in / there should be a clearer reference to that link i just put in here.

[Prime95]

Originally Posted by chalsall

For clarity, as you asking me to bring back into GPU72 the ~332.2M range for additional TF'ing, keeping the highly TF'ed candidates with Primenet for LL assignment? Or do you really mean 100M?

Easy to do either (or both) -- about 30 seconds of human work for the former; about 5 minutes for the latter. Just tell me what you'd like done.

I'm referring to the 332.2M exponents.

[Madpoo]

Originally Posted by Prime95

Watching prime95 do TF makes me very sad. We need to have GPU72 build up a reserve of 100M exponents that are already TFed.

Theoretically, could a massively cored CPU (like the current generation of Xeons) do decent TF work, if it were optimized? Or are we still in a situation where GPU's would still far outpace what a CPU like that can do?

I thought it was that GPU's were so much faster and people have already optimized code for it, so no attention was given at all to updating Prime95 to optimize TF for CPU use?

I'd imagine that TF work lends itself perfectly to multiple cores (P-1 and ECM as well).

While I'm wishing for things I'd also add in that Prime95 does stage 2 ECM work as fast as gmp-ecm.

With Xeon Phi stuff in particular, that's basically Intel diving into the GPU fray for the compute customers. When Knights Landing shows up in the CPU form factor, having Prime95 optimized for such things would be a benefit.

[Prime95]

Originally Posted by manfred4

There already is such a Job, but since that link is not available from the site itself, not many people use it.

The goal is to GPU72 use that web page. Until then it is also a good web page for telling you what TF ranges are in need of TF soon (i.e. will be handed out for LL shortly).

[chalsall]

Originally Posted by Prime95

I'm referring to the 332.2M exponents.

OK, yesterday I brought in everything from 332192831 to 332.6M TF'ed to less than 78 -- 3,562 candidates.

This leaves 22 candidates at 78 bits, 311 at 79 bits, 9 at 80 bits and 4 at 81 bits available for LL assignment from Primenet.

[aurashift]

Originally Posted by Madpoo

Theoretically, could a massively cored CPU (like the current generation of Xeons) do decent TF work, if it were optimized? Or are we still in a situation where GPU's would still far outpace what a CPU like that can do?

The fact that it only uses one core until it gets to P-1 or LL also makes me very sad.

[Mark Rose]

Originally Posted by Madpoo

Theoretically, could a massively cored CPU (like the current generation of Xeons) do decent TF work, if it were optimized? Or are we still in a situation where GPU's would still far outpace what a CPU like that can do?

It's basically a matter of flops. A GTX 580 running at stock clocks does 1581 GFLOPS. A four year newer 18 core E5-2699 v3 gives you 662 GFLOPS at base clock. Assuming the E5-2699 could be programmed as efficiently as mfaktc uses the GTX 580, you'd be looking at about 180-185 GHz-d/d, or 42% of the performance for 58% of the power usage.

I'd imagine that TF work lends itself perfectly to multiple cores (P-1 and ECM as well).

TF does. Doing TF might be an option for the "free" cores on memory bandwidth starved systems.

[Anonuser]

Originally Posted by Mark Rose

It's basically a matter of flops. A GTX 580 running at stock clocks does 1581 GFLOPS. A four year newer 18 core E5-2699 v3 gives you 662 GFLOPS at base clock.

I think that an E5-2699 v3 running at stock clocks does 2.3 • 18 • 32 = 1324.8 GFLOPS (single precision).

http://www.intel.com/content/dam/www...ions-paper.pdf