- May 3, 2016
- 31
- 0
- 18,530
My motherboard supports DDR3 1866 mhz, and it supports 64gb lrdimm modules.
I can find 32 gb lrdimm modules at 1866 mhz, or I can find 64 gb modules at 1600 mhz.
I can't find 64 gb modules at 1866 mhz, which leads me to beleive they might not exist.
My question is: does anyone know if they do exist, and if so how one might obtain them?
That's the tl;dr; version.
I feel the need to add that I'm aware this platform is old. However, with dual 12 core xeons it holds its own against my 16 core threadripper. In fact, it can outperform it while being much cheaper, given double the ram and the workload I'm throwing at them. I bought two for a reason.
With that out of the way, the difference in ram speed is more than academic. I'm using ramdisks, and the workload is ram throughput expensive.
With 64 gb dimms I could double the amount of parallel jobs, but I can't know if that will scale versus half the jobs using double the resources.
While 2 jobs scale ~40% better than one, I suspect a lot of that has to do with the fact that there are two numa nodes, and I account for them. Four jobs would mean fighting to share a numa node. Therefore I'm not even 100% certain 4 jobs are better than two with the same ram speed.
It would be a fairly expensive thing to be wrong about, given how much ram I'd need to test it.
I can find 32 gb lrdimm modules at 1866 mhz, or I can find 64 gb modules at 1600 mhz.
I can't find 64 gb modules at 1866 mhz, which leads me to beleive they might not exist.
My question is: does anyone know if they do exist, and if so how one might obtain them?
That's the tl;dr; version.
I feel the need to add that I'm aware this platform is old. However, with dual 12 core xeons it holds its own against my 16 core threadripper. In fact, it can outperform it while being much cheaper, given double the ram and the workload I'm throwing at them. I bought two for a reason.
With that out of the way, the difference in ram speed is more than academic. I'm using ramdisks, and the workload is ram throughput expensive.
With 64 gb dimms I could double the amount of parallel jobs, but I can't know if that will scale versus half the jobs using double the resources.
While 2 jobs scale ~40% better than one, I suspect a lot of that has to do with the fact that there are two numa nodes, and I account for them. Four jobs would mean fighting to share a numa node. Therefore I'm not even 100% certain 4 jobs are better than two with the same ram speed.
It would be a fairly expensive thing to be wrong about, given how much ram I'd need to test it.
Facebook
Twitter
Instagram