Topic: Custom RAM Timings for GPU's with GDDR5 - DOWNLOAD LINKS - UPDATED - page 5. (Read 155645 times)

you've not achieved undervolting in actual.  Check the load voltage during mining using HWInfo or gpu-z.  It would be much higher.

do you know, this will actualy not lower voltage on your memory, right? vmem is kind of bullshit.

Any ideas?

Undervolt made with SRBeditor.

Mem Controller:

300   @ 875mV
1000 @ 875mV
2100 @ 875mV

What could be wrong?

No way, something wrong there

Tried undervolting up to 875mv. On that value i got -10 watts @ wall. Not much... 

Thank you for your reply.

I will try to undervolt it by SRBeditor to 900mv. (Shouldn't be that much..) and let you know the readings on the wall.

You HAVE to undervolt...best test it through claymore, then if you really want it, put it in BIOS after
you find lowest stable voltage

EDIT: CALCULATION ON REAL POWER CONSUMPTION WAS WRONG. MY GAMING RIG CONSUMES IN IDLE 51 WATTS AT THE WALL SO THE REAL CONSUMPTION OF THE CARD IS AS HIGHLIGHTED BELOW.

my strap:

999000000000000022559D0010DE5B44A0551312B74C450A00400600750414206A8900A00200312 011112E36A42A3816

I've left voltage stock on bios. (means 1000mv on SRBeditor).
never used cvddc/mvddc on claymore , as i thought that reducing power consumption can be made by bios modifications.

Okay, let's go. What strap do you are using? Your hash is pretty good at 1200/2150

About power consumption: Do you are setting voltages manually? what cvddc/mvddc do you are using? RX570 often runs around 875~910mv cvddc at 1150. mvddc doesn't impact much power consumption.

Thank you for your specification.

At the moment with my RX570 sapphire nitro+ on hynix memory I'm reaching following hashrates:

1200 - 2150 Oscillating between 31.8 - 32.2. But with a power consumption that is far too high. (165 watts on wall).
1150 - 2050 Oscillating between 30.3 - 30.7. But with a power consumption that is still too high. (135 watts on wall).

I've talked to someone who shared a video proofing that on the same card with the same memory in a 6x rig he's making 31,3 stable for weeks and months with a consumption of 100-103watts per card. I'm hardly trying to reach that level.. but'm still far away.

Any hints on how to solve this?

Thank you in advance!

Straps doesn't change power consumption. Voltage/clocks do.

I'm reading into this for a while and wanna give to all the experts here big kudos for the huge amount of data that this thread produced. Thank you!

It took me a while to read through the 60 pages and lots of effort and even more trial and error to get some decent hashrate with the strap editor on SRB Editor. (btw, never used Excel so much ... to compare and edit the decoded straps).

My small but important request is:

I wanna find the sweetspot between power consumption and hashrate on, for example on one of the sapphire rx 570 4gb (hynix), but although I read (as far as I can.. there are many points that i do not understand.. on GDDR5 datasheets found through google) the decoded straps, i don't get it where to reduce significantly the power usage. (Sometimes it's quite frustrating )  Is it the strap in its complexity that makes the power usage or is it one specific point that makes the difference?

Thank you in advance!

Is the memory voltage in 560 locked?regardless of PBE stock value of 1000mv in reality it never sees anything above 850(overdriven seeing 850 and HWinfo 875 . meanwhile my 580 with hynix almost always locks to 950 forcibly whenever i 2100+ memory.
I'm certain its somewhere in the hex but not sure where to start. BTW same with all 560s i have 4gb and 2gb (got 4-5 different ones) one i'm using atm is GBrx 560 2gb OC gaming (6pin powered)
https://www.techpowerup.com/vgabios/192987/gigabyte-rx560-2048-170423

Hi all, i have succeeded to do relaxed timings from the one click ones etc (for bad hynix for example that would bsod with those one click too tight timings at even 1750 mhz memory)..... but GOOD memory is what i can't understand or explain, the phenomena for cryptonight being capped to maximum 540 h/s AND memory controller being at exactly 50% load (rx 560 2gb elpida) and thats it.... still the memory is a good batch and can clock more above 2000, but there is no increase in hash rate (deminishing returns for memory speed with one click elpida, starts at 1975-2025 mhz memory speed). Although i did the wise "other way around" of loosening timings alot, +2 tCL and tRP increasing tRC abit, tried substantially increasing RASMACTRD /WR (like that Hynix MJR strap with 55/66 respectively)and even tRFC to 150-160 value which is much for elpida straps.
Can someone explain this phenomena? What i thought i understand is that higher clock (as long as silicon/batch frequency allows) means higher hash rates but needs better custom timings to allow the higher clock.
What i also noticed is, the WIERD timing values i capture with gatelessgate, only some of the timings i cooked are there, rest are tottaly different or lower specially rasmact rd/wr the bus_turn group and ras2ras they are much lower live.

Appreciate any insight from the "Gurus"

It takes some time, but you can find most of the info in this and other threads on bitcointalk.  For example, you have to be very careful changing CAS timing, since one part of the strap contains the mode register settings that program the actual GDDR chips, and another part of the strap programs the memory controller CAS timings.  Wolf noted that you can usually get away with the two timings being one clock cycle off, but any more than that and you are pretty much guaranteed to get a crash.

Not a full dump, but some of the key strap timings:
https://github.com/nerdralph/amdmeminfo

@MoneroCrusher I know you already found my amdmeminfo fork, so this is more for other readers of this thread.

I am currently making a script to runtime mod the timing straps of AMD cards on Linux.
I can successfully overwrite timing straps while the GPU is mining. It only works if I use the same, or slightly abbreviated changes of *current* straps though.

Example of how it works:
1. GPU A is flashed with custom strap VBIOS
2. Miner is initiated for GPU A (to alleviate Mem Freq to get the desired timing strap activated)
3. Script is overwriting GPU A's Timing Straps with the ones it already has
4. GPU A continues mining

Example of how I want it to work (but for now it doesn't):
1. GPU B (exact same model as GPU A) fully stock
2. Miner is initiated for GPU B (to alleviate Mem Freq to get the desired timing strap activated)
3. Script is successfully overwriting GPU B's current timing straps with the one also used in GPU A
4. GPU Crashes

I mod the timing straps in this order:
I asked nerdralph if he has an idea and he told me timing straps are interdependent and therefore you have to runtime-mod them in a certain sequence for the GPU not to crash...
I was wondering if you could help me out with my problem. Would be really happy as I've wasted days of time now...

Probably not...PBE one click timings are in fact final result of hard work of people in this thread...you can get slightly better results
if you REALLY know what you're doing (that means detailed knowledge about ram timings and also detailed knowledge about
every crypto algorithm you plan to use)