Thanks. I forgot to mention I am running two per GPU right now. Given that, have others found the sweet spot on their hardware? I'm averaging about 540 seconds per blc26/guppy workunit using stock OpenCL_nvidia_SoG application.

Everyone else reckons that 2 at a time is best for such hardware, but every time I've tried it with my GTX 1070s & RTX 2060s there was no improvement, and when running different types of WUs (Arecibo & GBT) the run times for one of them wold always blow out.
It's one of those things you need to try for yourself & see how it goes.

Going one at a time with aggressive settings I'm doing similar WUs in 4min 30sec on my RTX 2060, 6min on my RTX 1070.

---

Grant
Darwin NT

I think people get so used to saying this is the set number that they forget all the other factors that go into figuring out how many units will run on a card.

Back when I was doing the testing of the SoG, I was running them on the 1080Tis . Those models have a higher amount of RAM compared to the 1080s and lower GPUs. Why is that important? Because of the way OpenCl runs on Nvidia Cards. Only 27% of a Nvidia card's RAM is available for scientific computations. Intel and ATI have a higher value of between 50-65%. So once you calculate what 27% is, it's easy to see why lower number of work units on lower cards.

The second reason is the type of memory. 1080Tis are running. It's GDDR5X as compared to the GDDR5 of the 1080 and lower. Better quality RAM and faster.

The last variable is CPU timing, RAM speed and the board itself. These are smaller variables as the first 2 above played the significant determining factor on how many work units I could get to run on a card. And since it's unlike we will be getting APs in large quantities anytime soon. 5 at a time on a GPU is the sweet spot for APs. If you have enough CPU threads to support them. The trade off is running a AP and MB causes the AP to slow down and MB to speed up.
But that is a mute point now a days.

---

It looks like running 3 per GPU has increased runtime to about 922-960 seconds for a similar blc26/guppy workunit. Additionally, according to GPU-Z, my clock speed increased from ~1410MHz to 1845MHz; temps increased to 70C; and power consumption increased from ~98W to ~150W. If I'm doing my math right, that's a 78% increase in runtime per workunit, with a 50% increase in workunits crunched.

If I'm doing my math right, that's a 78% increase in runtime per workunit, with a 50% increase in workunits crunched.

I always work it out on the Run time.
Roughly- For 2 WUs, the run time must be less than double the runtime for a single WU for it to be worthwhile. 3 WUs, run time must be less than triple for a single WU for it to be worthwhile.
Or work out the number of WUs per hour being done, which can be a bit of a headache to do at times.

For me 4min 30 for a BLC 25= approx 13.3 WU/s per hour (I can't find any completed BLC 26s in my task list at the moment).
For you 17min for 3 * BLC 25= approx. 10.6 WUs per hour.
(Edit- some of those WUs are done in 15min which would make for 12 per hour at that rate).

Command line values like mine even with just 1 WU would most likely bump up your GPU load & power use up significantly, as the defaults are set so as not to tax less capable systems.

---

Grant
Darwin NT

How many are you running at a time? What command line are you using?

Who is that addressed to?

---

My "in response to Message …" indicates it was to Grant, but I'm happy to hear additional input if anyone has some to share.

My "in response to Message …" indicates it was to Grant, but I'm happy to hear additional input if anyone has some to share.

I only run 1 GPU WU at a time. I have tried 2 at a time, with many different command line values, but have always come back to just running 1WU at a time on my hardware- GTX 1070s & a RTX 2060 (other than for AP, trying to run more than 1 WU at a time with SoG on my GTX 750Tis would be an exercise in futility).

---

Grant
Darwin NT