Unfortunately it's just down to trial and error to find the best combination for individual kit/board/clocks. Over time patterns can emerge for 'good' skews, but to make sense of it, you need to keep notes, change one thing at a time, and test and retest.
Found Memtest86+ very useful for DRAM skew testing and came up with this simple 'method':
If you run a deliberately low VDIMM, so that you get freezes in Test 3, you can see (quite) quickly which skews have the most benefit - for given clock and timings, certain skew combinations will pass multiple Test 3 runs, while others freeze instantly. I found the results were repeatable, and with the correct VDIMM (and VNB etc) applied later, memory was stable for full Memtest runs and then Prime etc, where it hadn't been before.
Trying, say, 25 skew combinations, I might get four which appear good and don't freeze - these tended to follow a pattern. The good skews all had the same offset between A and B, and differed only in the overall Advance or Delay from Normal (example: A2 DEL 25 / B2 Normal might work similarly to A2 DEL 75 / B2 DEL 50). At this point you then have a few 'good' skews to test properly and find which works best.
Don't know if this is really meaningful but it worked well for me. Adjusting CTL Skews gave me best results and I left all the others on AUTO.
Also, as setting DRAM SKEWS requires a full restart (IE your hard disks go off and on), I disconnect mine since Memtest86+ runs off a floppy or USB stick. That way you save your disks plenty of unnecessary spin ups
Reply With Quote
Originally Posted by cheapseats
Bookmarks