Topic: Custom RAM Timings for GPU's with GDDR5 - DOWNLOAD LINKS - UPDATED - page 6. (Read 155685 times)

I'm pretty new to this thread and although i've tried to do my homework, i can't seem to figure out if this hex method is more efficient than simply using the Polaris BIOS Editor (PBE) to mod my AMD RX series GPUs. Is there some sort of benefit to using this over the PBE, like will i be able to find more optimized timing straps leading to better mining efficiency?

Any inputs would be appreciated. Thank you.

Hi guys, started doing my homework and ingested alot of info from this forum and the Jedec Documents but need some professional help.
I have an 11 Rx 560 2gb since last January, i notice in bulk they are much more power efficient than 580/570 mainly in cryptonight.
Started by 2X Gigabyte one 6 pin powered and one not, sarcastically, BOTH clocked 2025+ with the Elpida Strap, they steadily do 540 h/s in CN V7 or old CN then bought many others, ALL of the others had a different Elpida batch "i.e lower bandwidth ones clocked at 1750, near thier maximum from the factory" thus, ALL of them can never go above 1900 Mhz memory giving ~500 h/s (one click elpida strap is good on low memory speeds)
FYI, many Elpidas BABG/Hynix AJR are of lower grade although they share the same name in software/bios (micron 4gb though, does have 1500mhz variants ending with "-60", which is at least honest).
Last GPUs i got were quite rare, Sapphire 2gb "6 pin powered" all of them Hynix AJR, these were a real disappointment, with one click strap they do 1850-1900 Mhz maximum, compared to my 5x 580/570 with a hynix AJR (clocking between 2025-2175) its Obvious here, that no silicon lottery here, thats only lower bandwidth stock Hynix AJR.

I now need help with:

*In general for all memory types except samsung  Cryptonight, is RRD=6 (5 being extreme?), Zeroing tFAW and t32AW of any worthiness or impact? what should i increase and which should i relax?

*trying to do a custom strap for Hynix AJR lower bandwidth for ~ 2000 mhz cryptonight.
I started with the one click strap, relaxing CAS latency, dropping the ACTRD back to 1625mhz strap levels and as much as i can try to keep TRFC and TRC within thier values, Relaxing TRCDR/A to 1625ish levels and decreasing TRCDW/A.
Do you think this is the right path for tweaking for cryptonight? I surely have to experiment, but my first experiment so far was wierd hash rate dropped, also i noticed some straps having RASMACTRD/RASMACTWR substantially relaxed. quite confusing here.

*Last, On higher quality elpidas, One click strap, nothing above 540 h/s is achievable, ~1975 mhz almost being the limit (well even on 1900mhz, it does 505 h/s) how/what path can i take to customize the strap for ~2050 mhz point.
P.S: 16 CU are alot/ increasing gpu frequency above 1150 mhz, yields no improvement whatsoever, so its not a compute limit issue, obviuolsly strap issue.
Anyway, appreciate any hints.
Edit:Forgot to add that I'm using (actually, not touching) the 1625mhz strap "learning?" parameters
MC_SEQ_MISC1 0x20140475
MC_SEQ_MISC3 0xA000896A
MC_SEQ_MISC8 0x20310002

But linked Hynix H5GQ2H24AFR (last seen in R9 290) is dated by 2009 and linux header is more recent (although if data is up to date here is questionable)

If that is so, it means data in .rom is mixed, so how can i know which data is in little endian, which one is in big ?
Or am i even getting this right ?

I really didn't knew about it. Awsum!

Do you know any place/app/mage/witch that decode these timings? It's something like "a new world showed for my eyes" lol

Again: I don't want a "ready" timings (okay, I would be happy with then) but just a way to do tests and tests and tests to improve performance.

does anyone know how to copy, mod or properly decode /sys/kernel/debug/dri/*cardnumber*/amdgpu_regs into readable hex code? when I try to copy it to another folder (to later view it in a windows hex editor) the kernel panics while reading the file.
would appreciate it a lot

I reset all GPUs in Overdrive before applying overdriven profiles in bat file.  

overdriveNTool.exe -r0 -r1
overdriveNTool.exe -p0"profile0" -p1"profile2"

Also all states except P7 on core and P2 on memory are disabled.

How do you fix your voltages? not a single windows app succeeds with me for Polaris.
i usually do as you said, 1225/2100 i mine cryptonight, always, the 560/570/580 take more voltage than the bios mod, then i try to give them 900-937mv in OverdriveN tool, never works, once miners auto start with PC, they get higher voltage (usually 950-1.1v) some garbage sapphire rx560D are 52% asic, these go to 1.1v almost! only when i click apply myself in overdriven OR use a VRM direct offset of lets say -8*6.25 (in supporting GPUs, most 560s doesnt have that).

Has anyone here experimented with RX 550?
With PBE I'm hitting up to 508 H/s (Powercolor 640 SPs), 505ish (Gigabyte 512 SPs), 500ish (Sapphire 640 SPs) and 485ish (Sapphire 512 SPs) all on Ubuntu Server.
Unfortunately tool support is pretty *wild* on Linux for the RX 550. For example amdmeminfo shows info about memories for all RX 550 with 512 SPs but doesn't return anything on RX 550 with 640 SPs. And for atiflash it's the exact opposite. It recognizes RX 550 with 640 SPs but doesn't recognize anything with 512 SPs.

So I have to use Windows for flashing.

Was wondering if anyone of you has played with it and could tell me how much more could be achieved with better straps?

Here are 2 memories that are used (only the ones with 512 SPs) that amdmeminfo could return:
Sapphire (EDW4032BABG)
Gigabyte (H5GC4H24AJR)

With recent drivers my 290x fall down to 820 H/S on CN vs 920 with the old drivers( although I do not longer mine with it, it was just a ponctual experiment).

I was doing 845h/s on my R9 290. BUT, that was at 1100/1500. And with wolfminer, with bad configuration. From what i have done, it seems like the card is heavily core-bottlenecked. Modifying the memory straps yielded 0 performance increase. Overclocking the core brought about 100h/s from stock. This was also pre-blockchain drivers. I would suppose it can do 1kh/ and a tad more.

Any of you guys still using any old legacy GPUs like the Pitcairns, Tahitis or the Hawaii's?



Hawaii's are still good with ETH however very power intensive. Anyone manage to get more than 800h/s with them on CN?

Pitcairns are basically best for Cryptonite algo if they got Samsung or Hynix memory. Can get like 520H/s on some of the R7 370 GPUs.

Then there is the best mining GPU of all time. The killer Tahitis. They dominated mining BTC back in 2012, then LTC back in 2013-2014, then Darkcoin in 2015, then ETH in 2016, then ZEC in 1017.

Now you got 2 options with them. Either Equihash or Cryptonight.

Equihash is around 300H/s however it requires almost 200 Watts

CN is around 600hs and only uses 150 Watts. Depending on ASIC quality some can run at the lowest voltage possible 0.95V and consume same amount of power as a Polaris GPU while hashing at 550h/s.

Wondering if anyone can squeeze more than 600h/s out of these old monsters? Maybe by memory strap mods.

I merit whenever I get some smerit.  i don't know how often I get smerit either. :/

Well you guys have way more experience than I do with this stuff but these were the only settings not fully explained in any guide or discussion. However they seem to be different between cards with same memory. They also seem to change between some straps like say 1:1300 to 1:2000.

So it sounds like when people are tuning their straps they do not touch these just leave them as they were?

Those are dram training timings, they affect a lot if you know how to tune them, they contain DRAM clock latency, write recovery, write latency and RDBI/WDBI training bits + a lot more that usually won't touch.

Those are the tetriary (or are they tetriary?) timings. They do not affect have much effect on sped, a bit on stability. They do not weight much at all. I have copy/pasted them most of the time. They include the stuff like Memory Training, Power-up sequence, Mode registers blah blah. At least thats the impression i have been left with.

How are these items set. I noticed that they are changing between timing straps for the same Memory but I am unsure what they are dependent on or correlated to. Are they Tied the physical card or are they determined by settings in the CAS & RAS timing. If you have a search term to use I will go look it up on google, however when searching these terms I get this forum and other Memory Strap guides but no explanation on how to properly set them.
####MC_SEQ_MISC####
MC_SEQ_MISC1 = 0x2014080C
MC_SEQ_MISC3 = 0xAB0089EA
MC_SEQ_MISC8 = 0x00000003

And why would I do that?

P.P. to the point... as VirosaGITS said G5 is already pretty optimized, I was able to gain additional 0.5MH from samsung on p106, but for micron it was pushed further, from 22.5 to 24.7, though with some memory errors.
What OhGodACompany is doing is by all means fast profit due to upcoming GPU obsoletion. I asked OhGodAGirl several times on IRC but got no response, so I'll ask here - Why now?

Laik2, please send me PM, i would like to send you some text, but disabled.

Thank you

Cryptonight is affected a lot by the TRC and ACT/RASMACT timings. Just dont go too tight on those, aside of stability issues you will start running into negative performance gains.