Post
Topic
Board Mining (Altcoins)
Merits 4 from 2 users
Re: AMD Memory Tweak - Read and modify memory timings on the fly - [Vega Friendly]
by
pbfarmer
on 25/03/2019, 18:59:46 UTC
⭐ Merited by frodocooper (3) ,vapourminer (1)
I played with this tool during the weekend on Vega.
But to be honest I did not see any hashrate increase under CN algos.
I'm not an expert on timings, but I should see at least 1-2% increase when changing CL 20->19

That was the first time I put a Linux AMD system together and I am struggling with the different kernel and driver versions not working together.
The biggest problem is the GPU freq is affected by the voltage.
If I set in PPT P7 1600 950 then the GPU runs only at around 1400@950mV. If I increase the voltage the clock goes up, but quickly reaches TDP.
Also if I try to trick it and set P7 1700@950 then if load changes freq spikes up and card freezes obviously.
Tried with 3 different kernels and one is worse than the other.
Maybe I should try other than the 18.50 amdgpupro?

Ubuntu 18.04 + amdgpu-pro 18.5 is what I use, without problems.  The voltage-frequency relationship is a Vega feature (ACG), not the driver.  Same behavior is observable in Windows, though not as aggressive for some reason.  Every card is different, but depending on the algo, a 1500 cclock setting in Windows (to get to 1400 effective for example) may require a 1600-1700 setting in Linux for the same result (@ same voltage).  The trade off is while in Windows a too aggressive cclock overshoot generally caused a crash/hang on load transitions, I do not see the same problems on Linux.

Also, fyi a ~1400 effective cclock is generally achievable with <= 850mv.  950 is really high for almost any algo, unless you’re running mem above 1107 (hint: don’t - it’s not worth it wrt efficiency.)