All things Radeon VII / Vega 20

I updated to driver 19.4.3.  That seems to have given me a slight elapsed time reduction at 2X--but maybe that is some other side effect of rebooting.  Then I tried switching to 3X.  So far I have zero compute errors in over five hours of running.  Last try I had accumulated ten Computation error 65 instances by then.  Something is different.  Maybe it is the driver, maybe something else.  So far the productivity gain is very small at 3X versus recent 2X--but for the first five hours, I was running with the CPU affinity set to just two cores.  I'll probably get at a slight further boost from having just now loosened it to three cores.

My 3X adventures are over. and I have resumed running 2X.  I have these conclusions:

1. Driver 19.4.3 gave me a slight 2X productivity improvement over 19.3.1 on Einstein GRP work.
2. My system as currently configured gives an unacceptable rate of "error while computing" type 65 when running 3X.
3. The 3X error 65 rate is substantially raised by changing from two to three allowed cores for CPU support (as constrained by affinity)
4. The 3X error rate seems least when I am not myself doing things on the system, and raised when I use browsers, editors, and such.
5. Relaxing the power limit from the starting point of -16% as far up as -9% in 1% increments did not find a level at which the error while computing problem disappeared even with CPU support restricted by affinity to two cores.

While it might be that going to zero power limit would finally find acceptable 3X error rate (for this type of error I currently consider the acceptable rate to be zero), I would not run there as the extra power consumption troubles me more than the modest productivity improvement delights me.

I doubt there is anything fundamental about my results.  Other samples of the Radeon VII card running on other systems (let alone running other applications) may well do just fine at 3X.

Interesting.  This source advocates power reduction by turning the core voltage knob, not the power limit knob.  My personal observations on my card in my system on Einstein work has been bad behavior in response to imposition of reduced core voltage using that knob, and much better behavior in response to the power limit knob.

The real mechanism of power reduction must primarily depend on the core voltage and core clock rate, regardless of which knob one turns to get the result.  But the card internal controls continue to make very frequent adjustments of both voltage and frequency.  The key question is what user input configuration causes one's card to make the best choices.

My personal Radeon VII current configuration involves control using Afterburner (I like the smooth fan speed I could get out of it much better than the motorboating I frequently got under Wattman control) with the following user settings:

Core voltage -- default (reads max 1123)
Power Limit -- -16%
Core clock -- default (reads max 1801)
Memory clock -- default (reads max 1000)
Fan speed is on a user map

As to the average actual operating condition, GPU-Z reports these averages:

GPU clock: 1605 MHz
Memory clock: 1000 MHz
GPU voltage: 1.07 V
GPU only Power draw: 207 W
GPU Temperature: 79C
GPU Temperature (hot spot): 98C

I've seen multiple reports of people seeing quite severe errors when adjusting down the voltage limit by surprisingly small amounts on the Radeon VII.  I speculate that the current cards are shipping with controls which are not very smart about setting the other parameters to workable levels when user voltage limitation is dialed down.

Quite likely these matters are very workload dependent.