Haven't really done much in the past w/ CN light variants, as I got the impression they were a bit more power hungry... Is that the case w/ this (e.g. compared to cnv4)? Just so I can get a sense of where to start my voltage settings.
Not trivial to measure, there are a few different trade-offs in play, but I'd actually say they're about on par. CN-turtle uses the CNv8 main loop which (in our case) draws less power than the CN/r main loop, but CN-turtle does for sure put more load in general on the gpu than the 2MB pad variants.
Ok, only Vega 64 for now, and only from a single GPU (though voltage needs should really be the only variable moving forward w/ addl GPUs). Will try to post VII and 590 numbers later:
Env: ubuntu 18.04 + amdgpu-pro 18.50.
Best cn_config - based on initial test: L28+28 (see table below)
Test desc | |GPU (actual cclock, mclock, mv) | |Miner h/r (kh/s) | |Power ATW |
|CNR settings | |1390 MHz, 1107 MHz, 825 mv | |19.60 | |? |
|mclock+ | |1390 MHz, 1145 MHz, 900 mv (not tuned) | |20.18 | |? |
|mclock+, cclock-- | |1210 MHz, 1145 MHz, 862 mv | |20.14 | |~170w |
|Final settings | |1210 MHz, 1107 MHz, 806 mv | |19.54 | |~150w |
Notes:
- Definitely lower power requirements vs cnv4/cnr if tuned properly
- Algo performance is heavily dependent on mem bandwidth, not so much on core clock speed.
- Increasing mclock from 1107 to 1145 gave ~+3% h/r, but at a heavy cost (+13% power) due to SOC scaling to 1200.
- Dropping core clock from ~1400MHz to ~1200MHz only reduced h/r by 0.25% (19.60 to 19.54 kh/s) - all the way down to ~1100MHz still gave 19.45 kh/s, but 850Mhz resulted in ~12 kh/s, so there is a cliff somewhere, or maybe I needed to re-adjust cn_config
- I didn't do any more tests reducing cclock further, as I was already down to the voltage floor at 806mv/1200MHz. However, if we can go significantly lower (like ethash,) we may be able to save enough power in cclock reduction to make an mclock boost worth it???
- Finally another algo that would seem to make the 56->64 flash worth it.
- cn_config 'L' variants not only had better h/rs, but seemed to also put less pressure on GPU, therefore requiring slightly less power
- Windows 10 pro w/ 18.11.x drivers produced exact same results at my efficiency settings in linux, so didn't bother testing further. In general, I find w/ most algos that as long as the effective (under load) clocks match, h/rs and power use will match as well. The only real diff between linux and windows is the cclock setting required (much higher) to get to your desired effective cclock.
cn_config | hashrate | ___ | ___ | cn_config | hashrate |
L28+28 | 19.60 | | | 28+28 | 19.32 |
L28+30 | 19.57 | | | 22+24 | 19.32 |
L26+28 | 19.57 | | | 24-24 | 19.32 |
L26+26 | 19.56 | | | 20+20 | 19.30 |
L24+24 | 19.52 | | | 23+25 | 19.08 |
L20+20 | 19.48 | | | 25+25 | 19.06 |
24+24 | 19.44 | | | L30+30 | 19.06 |
L28-28 | 19.39 | | | 22+22 | 19.01 |
24+26 | 19.36 | | | 23+23 | 18.95 |
26+26 | 19.34 | | | L27+29 | 18.89 |
26+28 | 19.33 |
I'm gonna give my 56's more of a test run later. I run them at XMR clocks currently - which means I'm feeding them unnecessary core clocks and voltages. But even then the cards aren't sucking much power. I'll report later... when I get some time to test