Niall’s virtual diary archives – Friday 30 August 2019

Friday 30 August 2019: 12:54. Last post I mentioned that I had plumped for 8x M391A2K43BB1-CTD Samsung 16GB DDR4-2666 ECC UDIMM PC4-21300V-E Dual Rank x8 Module sticks in order to make 128Gb of RAM for my AMD Threadripper 2950X workstation with an Asrock X399 Taichi motherboard. These are not on the QVL list for any X399 motherboard, so it was a shot in the dark. But for €564 ex VAT they were a steal, and one or two reports on the internet said they could be overclocked easily to 2933 at tight timings, after which the Threadripper supposedly won’t go faster due to the 8x RAM sticks in any case, so buying faster RAM seemed pointless, and the ECC and Samsung chips seemed a great bet for the money.

I can report no disappointment. One of the sticks was faulty, so I cannot usefully report bandwidth, but these 2666 sticks will happily do DDR4-2933 @ 1.35v 16-15-15-36-50 by simply keying in the values in the BIOS. I haven’t changed anything else from stock.

I did try CAS 14, but it won’t POST. Internet says that Threadripper rounds odd numbered CAS up to even in any case, so that report of CAS 15 working I mentioned last post he was probably at CAS 16 and didn’t realise.

2933 @ CAS 16 = 10.9 ns. My previous GSkill RAM was 3200 @ 16-18-18-38-56 = 10 ns. Benchmarking the difference in Windows, this overclocked Samsung RAM is now exactly on par with default configuration Threadripper v2 @ 71.2 ns main memory latency, whereas the previous faster GSkill stuff benchmarked at 67.7 ns, which is a 5.2% increase in latency from the CPU.

How much of that is due to the main memory being slower, or simply the Threadripper’s internal fabric running 9% slower, I cannot say. I would say that the non-CAS timings are much tighter for the Samsung memory, so maybe that explains why the latency increase from the CPU is half as bad as the internal fabric and RAM would suggest.

Anyway, all in all everything seems to be working fine with this 128Gb temporarily minus 16Gb setup, and the ECC is operating correctly according to Windows (no indication in the BIOS though). Just need the defective stick to be replaced, and I’ll be able to tell you about effects on memory bandwidth.

#threadripper