its now running at 3.8 ghz...
Let's see:So, @StefanR5R , how is 18 hours on this project ??
There is another level of memory tweaking beyond voltage you can do with sub timings. I was able to drop the latency on my memory kit from like 98ns stock to 64ns with fairly minimal fuss. I can grab my timings if you would like to play with tweaking at that level, but I don’t know much about MSI boards to say where to find the settings. I also don’t know if any of test tools like aida exist for Linux, or if you would mostly just be testing by running prime grid.Well, before I get pissed, I will wait until these units finish, and see how the next does (like yours). But at least I know the memory is finally tweaked and working.
well, if you could pass those along (the sub-timings) as I don't know any linux tools to do that.There is another level of memory tweaking beyond voltage you can do with sub timings. I was able to drop the latency on my memory kit from like 98ns stock to 64ns with fairly minimal fuss. I can grab my timings if you would like to play with tweaking at that level, but I don’t know much about MSI boards to say where to find the settings. I also don’t know if any of test tools like aida exist for Linux, or if you would mostly just be testing by running prime grid.
i also don’t know what the gains are in regards to prime grid. Obviously if it doesn’t hit memory very hard, tweaking sub timings won’t help much if at all.
PrimeGrid's LLR based subprojects generally hit the RAM rather hard. On processors without enough last level cache (currently for each SoB-LLR instance: ~25 MB, contiguously), this workload is memory-bound and leaves the execution units, notably the FMAs, rather underused. On processors with enough cache, e.g. Ryzen 5950X if one SoB-LLR instance is executed within one CCX, it appears there is still unusually heavy RAM access. (Remember the crashes which Mark reported initially.) But the SoB-LLR performance will definitely depend considerably less on RAM performance than for processors with smaller cache.i also don’t know what the gains are in regards to prime grid. Obviously if it doesn’t hit memory very hard, tweaking sub timings won’t help much if at all.
BTW, the initial crashes were memory running at 1.22v instead of 1.45.PrimeGrid's LLR based subprojects generally hit the RAM rather hard. On processors without enough last level cache (currently for each SoB-LLR instance: ~25 MB, contiguously), this workload is memory-bound and leaves the execution units, notably the FMAs, rather underused. On processors with enough cache, e.g. Ryzen 5950X if one SoB-LLR instance is executed within one CCX, it appears there is still unusually heavy RAM access. (Remember the crashes which Mark reported initially.) But the SoB-LLR performance will definitely depend considerably less on RAM performance than for processors with smaller cache.
However, I don't know how much RAM latencies matter, versus RAM bandwidth.
Incidentally, I tried llrSOB on Epyc 7452 (32-core Rome) with only 6 memory channels populated. Throughput fell into a bottomless pit. It's less than half compared to a full octa-channel configuration.Forrest Norrod said:In Rome or in Naples, with eight channels of memory you could get full performance - you could get a pretty good well optimized and balanced system with only four channels of memory, obviously your theoretical bandwidth is cut in half, but it was well optimized. If you had six channels of memory, you get this somewhat unbalanced condition where latency and throughput [would depend on a number of factors], so that’s what we’ve really tried to address with the six channel to give that additional flexibility to right size the amount of memory for your workload without giving up any of the performance.
Its at 8.3% done in 12 1/2 hours. Not good.At least on Linux, there is no need to disable SMT. Just tell BOINC to use 50 % of the logical CPUs.
In the AMI BIOS of Supermicro boards, the SMT setting is supposed to be in "Advanced" -> "CPU Configuration".
Boincmgr's estimation of time remaining is not reliable. Perhaps better is to check how fast (well, slow) the completion percentage is progressing.
This one is probably as close as any as it pertains to this specific challenge - https://forums.anandtech.com/threads/primegrid-challenges-2021.2589211/post-40462572Is there a post in this thread that I should follow for configuring Primegrid? Sounds like just going with whatever it configures on its own isn't a great idea?