Post
Topic
Board Mining (Altcoins)
Re: AMD Mem Tweak - Read/modify mem timings on the fly - [Vega Friendly] [Win/Linux]
by
pbfarmer
on 26/05/2019, 04:54:49 UTC
For anyone else using the old PBE one-click samsung straps (aka generaleramon's ubermix 3.2) on RX 4/5XX 8gb GPUs, and seeing subpar ETH h/rs due to the tight timings limiting mem clock frequencies... I've been able to get my samsung 580-8s more in-line w/ the microns (~32.5 MH/s) by loosening a couple settings which then allow raising the mem clock (combined w/ a REF increase i use on all my 580s):

--REF 10 --RC 70 --RFC 157 (from REF 5, RC 65, RFC 136)
mclock 2145-2185 (from 2085-2145)

other relevant bits:

cclock 1175, 806mv, phoenixminer, and ubuntu 18.04 + amdgpu-pro 18.50 (though windows should behave similarly)

You could prob get to ~33+ w/ a 1200-1250+ cclock, while I may have slightly overshot on the loosened timings.  But as usual, I'm more interested in efficiency than max h/rs.
Also, mclock is still a bit limited, as i believe CL (currently 21) would need to be relaxed to get to 2200+ MHz -- which as i understand, can't be done w/ the tool atm.

My Samsung memory is weak, only able to run at 1960 with Ubermix.  I was able to bump to 2000 with your suggestion along with a REF 30 and gained 0.7MHs.  I would definitely have to loosen Column Address Strobe Latency to go higher.

That's interesting - I had always thought I was limited to ~2K on ubermix as well, but during this exercise of finding a way to get to higher clocks, I found out I could get to ~2.1K w/ no timings changes.  I guess the 2K limit (at least for me) was more the case under CN - ethash must be a bit easier on the mem.

I'm also surprised you can do REF 30 - I can do higher refs w/ CN, but for ethash, I'm seeing a couple percent errors already at REF 10 (def depends on ambient tho.)

Maybe we're using different ubermix versions?  iirc, a couple different ones were included in PBE over time.  I also recently noticed generaleramon had published later versions (even past 3.2,) one of which was suggested may be ideal for 2150.