|Moar performance testing
||[Wednesday 8th November 2017 at 11:57 pm]
So as part of the grand performance test experiment I spotted that my PC wasn't set up as best as it could be. In particular, the RAM I've got is certified with better timings than it's actually running with. The only reason it's not using them I think is down to the particular chipset/BIOS I've got (as well as the JEDEC profiles the SPD chip also contains an EPP profile but that appears to only be understood by nForce chipsets). I'm also not convinced the RAM was properly running in dual-channel mode as CPU-Z listed it as being Dual/Asymmetric.
A BIOS update solved the dual-channel configuration - it turns out I was running a rather early version of the BIOS for this board - and I re-ran PerformanceTest to get a fresh baseline. Unlike the original test I also ran this after a clean boot with only background tasks active, and I used RamMap to page out as much as possible (because PerformanceTest considers free physical RAM when calculating the memory score - which is bonkers, especially when having a SSD means lack of physical memory really isn't an issue even for memory-hungry programs).
Clean run, BIOS 1305
The result is a nice increase from 1619 to 1791, with improvement to everything bar disk (possibly that suffered from contention when paging stuff back in?). This is with the auto-detected RAM timings of 5-5-5-18 @ 400MHz 1.8V.
Next is to adjust the RAM timings. This was fairly straightforward - just drop into the BIOS, enable RAM overclocking and update the timings (leaving everything else on Auto). I also had to boost the voltage to 2.1V - to do that required setting Ai Overclocking to Manual and then overriding the DRAM voltage. And the result?
Clean run, BIOS 1305, RAM timings 4-4-4-12 @ 400MHz 2.1V
I'll take that 27 point increase, thankyou very much. I never expected a significant change and there's enough variation in the other numbers to largely hide this in the noise. It becomes a bit more interesting if I look specifically at the memory score:
Memory Mark with 5-5-5-18 @ 400MHz 1.8V
Memory Mark with 4-4-4-12 @ 400MHz 2.1V
That's nearly a 10% reduction in latency and a corresponding substantial increase in sustained memory bandwidth, and all technically without overclocking my system (although pleaseremove disputes the lack-of-overclocking, given that I've run Nyx with the slower timings for the best part of 10 years). Nice!