I’ve got a Threadripper 2950x running in an Asrock x399m taichi motherboard, with 64 gb of Corsair Vengeance LPX 3000 MHz RAM (4x CMK16GX4M1B3000C15), which I use a machine learning workstation. I’m running on the latest BIOS (3.7.0). This is my first desktop build and I don’t have any overclocking experience.
Unfortunately, my RAM is running at the base speed of 2133 MHz instead of it’s rated speed. I first tried just checking the enable XMP button, and windows blue screened before booting. I walked the XMP rating down in steps all the way down to 2133 (but with tighter timings), and even though I was able to get things to boot, I ended up crashing again after a few hours of heavy load.
When I look at the Motherboard’s QVL list it has several 2933/3000 C16 Corsair memory modules qualified at 2933, so it seems like I should be able to get my 3000 C15 modules to run at 2933, but I haven’t been able to do so.
This morning, I decided to get more serious about dialing in an overclock manually, and I started following the MemTestHelper guide, which suggested starting by backing off on the timings (to 16-20-20-40), and then seeing how high of a clock you could get. I was able to get booted at as high as 3000 MHz, but it crashed in a matter of minutes. I started running memory stress tests with Karhu RAM Test, and even with the relaxed timings, I was getting errors down as low as 2400 MHz.
I poked around on the level one forums, and I found the “Memory Unleashed on Threadripper: 128gb & 2933 & ECC tested” thread about running threadripper with 128 gb of ram, using G.Skill 2933 modules. There was a post suggesting that I increase my SOC voltage to 1.1 V, but even that + relaxed timings doesn’t let me run at 2933. Also in that thread it was suggested that the G.Skill modules used are the best modules available for threadripper, and I looked into buying some, but it looks like a full set of 4 would cost 850 bucks, which doesn’t seem worthwhile.
I’m sort of tearing my hair out, since with the few minutes of benchmarks I’ve been able to run say that if I can get 2933 stable, I’d be looking at 16% performance uplift (machine learning with tensorflow), which would be really nice. I’m still running tests to see if I can get any overclock stable with 1.1 V SOC + relaxed timings. I do need an incredibly stable overclock, since ML training runs can take multiple days to complete.
Is there anything obvious I’m missing? I’m a little concerned that I ended up with memory modules that are too low end to really get any sort of overclock, since the ones I have are 1/4 the price of the “recommended” G.Skill ones - is that correct?