Originally Posted by snurds
So it is possible to get a benefit at stock RAM clocks by copying the lower timings to the higher straps?
Why does this work? Did AMD make an error of some sort which is reflected in the irregular progression of straps from 1375MHz to 1731MHz then back to 1500MHz?
Just trying to understand it a little better.
Let me share my results with everyone interested and possibly help you a bit Introduction, just for better understanding (Click to show)
As you may know, some time ago I've performed the RAM replacement procedure on R290X and changed the RAM chips for Elpida 4032BABG. It's not a very popular IC type, there's only one Elpida VBIOS at TPU database (from Sapphire R390) and a couple of VBIOSes from R290X 8GB. VRAM_Info table in Sapphire's BIOS contained 11 straps (200-400-800-1000-1250-1375-1425-1500-1625-1750-2000), while both 290X 8GB had only 6 (200-400-800-1000-1250-1375). Not too much for the beginning...
After that I tried to search for another Radeon cards that might have been equipped with the same RAM. I searched among 7790, R260X and R270X. I successfully found the 12'th unique strap for 900MHz. It had one improper sub-timing, possibly because of another GPU, but I modified it using simple calculations and some approximation.
So, in the end, my full set of 12 straps suitable for both 290X or 390X /w Elpida 4032BABG was ready.
12 straps for Elpida 4032BABG (Click to show)
200M: 204E00 00 999133200000000060881107C0540B060F05C1000020410022AA1C08150014209A8840A1000000C0030105070C0A100C
400M: 409C00 00 99913320000000006094120FD0540C07158892010020410022AA1C081D0314209A8880A2000000C006010A0F190E160D
800M: 803801 00 9991332000000000A5AC351F10550E0B218E35030044820022AA1C083D0914202A8900A5000000C00C06141A33182210
900M: 905F01 00 9991332000000000E7B4362430550E0D2590A6030044820022AA1C084D0B14202A8980A5000000C00E08171D3A1C2611
1000M: A08601 00 77713320000000002939572750550D0E261107040068C20022AA1C08540C1420AA8900A6000000C00F0A191E401E2712
1250M: 48E801 00 7771332000000000AD49593270550E102D1519050068C30022AA1C08640F1420BA8980A7000000C0130E202551242E13
1375M: 1C1902 00 7771332000000000EF516A3790550F1232179A05006AE40022AA1C0874021420CA89C0A8020000C01510232859283315
1425M: A42C02 00 777133200000000010D66A3990550F123498CA05006AE40022AA1C0874031420CA8900A9020000C0161124295C293515
1500M: F04902 00 7771332000000000315A6B3CA0550F1336191B06006AE40022AA1C087C041420CA8980A9020000C01712262B612B3715
1625M: C47A02 00 777133200000000073627C41B05510143A1B9C06006C060122AA1C0804061420EA8940AA030000C01914292E692E3B16
1750M: 98AB02 00 7771332000000000B56A7D46C05510153E1D1D07006C070122AA1C080C081420FA8900AB030000C01B162C3171313F17
2000M: 400D03 00 999133200000000018F77E4F0054121A46A01E08006C070122AA1C08350C1420FA8980AC030000C01E1932378139471A
note: there're two excessive tabs in each strap, after 6th and 7th bytes, for easier viewing, don't forget to remove them before copying the values into hex editor
Now it was the time for overclocking and benchmarks. I used the MSI Kombustor GPU Memory burner test as it seems to be quite sensitive to memory speed and latency. And certainly, HWiNFO for EDC reporting.
At first I wasn't satisfied with the results, the RAM could pass the 1500MHz barrier, but it began to produce errors at ~1530..1540MHz, but even at 1500MHz it wasn't rock stable. The max error-free speed was about 1450-1480MHz.
I tried to copy the timings from the Nth strap into the (N+1)th zone, but I still observed performance drop, this method didn't work...
So, the only thing I could do is to increase the Nth strap "length", increasing the (N+1)th starting frequency value.
...Many hours have been spent during testing... And what's in the end?
Now the RAM runs 800 strap up to 1250MHz(!), 900 strap up to 1350MHz(!), and 1000 strap up to 1550MHz(!!)
Moreover, trying to lower the VDDCI I came to 0.925V, and it's still not the limit, and the RAM is rock stable @0.925V and 1550MHz with 1000 strap (!!)
So, I still can't believe it, but it's true. And Elpida 4032 is not as bad as I initially expected
Here's my current VRAM_Info table, the straps after 1550MHz are not critical to operation so I set them nearly randomly
But the critical straps are okay.
VRAM_Info_4032BABG.zip 2k .zip file