Prime95 Work Preference
 

What do you guys let your CPU do? I have a 4 core / 4 thread 3.4 GHz that's somewhat ok at trial factoring at the moment, but my GPU can do the trial factoring so much faster that it seems like a waste to be running it on CPU. I was doing elliptic curves, but when the exponents are in the 9m range it starts taking an extremely long time (and you only get about 1 GHz-d, so I figure that means they want to discourage us from ECM for now?). I have a Mac running trial factoring to low limits that amusingly gets about 11.5 GHz-d/day, which I'm fine with. But should I be using my desktop CPU for something like double check LL's? Or P-1? Or what?

[QUOTE=tapion64;371169]What do you guys let your CPU do? I have a 4 core / 4 thread 3.4 GHz that's somewhat ok at trial factoring at the moment, but my GPU can do the trial factoring so much faster that it seems like a waste to be running it on CPU. I was doing elliptic curves, but when the exponents are in the 9m range it starts taking an extremely long time (and you only get about 1 GHz-d, so I figure that means they want to discourage us from ECM for now?). I have a Mac running trial factoring to low limits that amusingly gets about 11.5 GHz-d/day, which I'm fine with. But should I be using my desktop CPU for something like double check LL's? Or P-1? Or what?[/QUOTE]


Double checking LL is much faster on CudaLucas, P1 work can now be done with CudaP1. (some of those tools are forever in Beta)

However I do ONLY first time LL tests on my CPU, unless is needs P1 first which just automatically gets done in the process of the LL test.

My opinion is 4 core @ 3.4ghz is best served doing first time LL testing.

Others will have different answers, opinions, suggestions.

At the moment I'm running 4 LL first tests on CPU, and running one of those on GPU at same time. All 4 LL tests are set to finish around June 24th, giving roughly 70 days. The GPU will finish one of them in around 7 days, so I could go through all 4 of them and get a result before the CPU is done. If GPU comes back negative on any of them, should I stop CPU? Or let it continue and consider it a double check?

I used to dedicate my 4 cores at home and 2 at work exclusively to p-1 work, but now that is not as critical and fast on the video cards. I pretty much only do first time LLs now on the cpu at home, but I still do p-1 at work just because if I found a prime at work it'd cause problems.

EDIT: You didn't exactly ask about what to do with your gpu, but the gpus are a lot more efficient with the TF work than the LL stuff. You could dedicate all the cores on your cpu to one LL test if how fast they got done was your only measure, but it's more efficient to run one LL per core. Similarly, running all 800 gpu cores against a single LL will get it done fast, but it probably still lends more use to the project to TF or p-1 on the GPU.

Well, I figure a week to do a first time test isn't that bad, might as well tackle a few. I'm going to pick up an r9 295x2 once it becomes obvious which non-reference card is the best, and that'll more than take care of my TF contribution. I've also been doing GPU72 trial factoring up to 71 for a bit, but I wanted to switch it up.

If you have a lot of RAM and aren't afraid to put a lot of it to work, you could run one P-1 worker with as much RAM as you're willing to spare and put the other three to LL or DC. My CPU is currently unable to work but it is assigned two LL and two DC, just because I feel DC is completely neglected.

TF is a clear winner for GPUs. Yes, LL tests are faster but you will find that they're not so much faster that they're worth it. They get maybe 20% as much GHz-Days doing LL as they would doing TF, and TF is where we're missing the most firepower at the moment.

Of course, you are always free to do whatever you want with your hardware. ECM isn't discouraged. It's just more of a sub-project than anything.

I have an FX-8350 doing DC on all cores: roughly 2 per month per core.

Hmmmm. Ok. I have 12 GB ram that I'm almost never using, even when I was doing 4 ECM on 9M exponents I wasn't running out of memory during stage 2. I guess after these LL tests finishes on CPU I'll switch to 1 P-1, 1 first time test, 2 double check tests. Then 560 Ti will stay on LL first time, and the R9 295X2 will be a trial factoring beast.

[QUOTE=tapion64;371262] Then 560 Ti will stay on LL first time,[/QUOTE]

Are you sure the GPU is reliable at LL testing? GPUs are sometimes stable and reliable at TF, but give out errors at LL testing (that happens with my 560Ti). It would probably be a good idea to start by running some DCs on the GPU and verify they match the first time LLs, to be on the safe side of things. If and when you get confident it is reliable, change to first time LLs.

Depending on brand and model of card, memory errors can show up when running CUDALucas, and not be a problem with mfaktc. I could not complete the CuLu self-test on a Gigabyte GTX570 until I reduced the memory frequency by 200-300 MHz, i.e., from 1900 down to 1600 or 1700.

I agree: Stick with DC or TF for the easiest time. DC's are shorter, so there's less time for something to go wrong, and more results to check with.

These are primarily gaming cards which don't need their memory to be absolutely spot-on perfect, but CudaLucas definitely does, so you have to underclock the memory.