gpuOwl Windows setup for Radeon VII

[Prime95]

mfakto ran 280000+ self tests without issue.

[Prime95]

With mfakto TF running in one window generating nice stable core and memory clock speed I tried gpuowl prp on 216091. No joy, ending the power piking theory.

[tServo]

Originally Posted by Prime95

add "-use ORIG_X2" to config.txt in the gpuowl directory

Thanks,George.
That fixed the __asm problem but now I am having the same results as you, even with the Afterburner stuff I modify. I get a black screen and must power cycle.
I have also tried it without -fft +3 to no avail.
Maybe more experimenting tomorrow …..
I might try underclocking it.

One thing I did notice, when starting v5.0, which still works of course, Afterburner shows
a 10 degrees C jump upon startup INSTANTLY.

[kriesel]

Originally Posted by tServo

One thing I did notice, when starting v5.0, which still works of course, Afterburner shows
a 10 degrees C jump upon startup INSTANTLY.

Gave 5.0-9c13870 a try here, got encouraged by the smaller exponents, then it hung on the start of 6972593.

[Prime95]

I ran a two hour mfakto run without problem.

I tried some suggestions from here, namely eliminating game bar. In Wattman I increased the performance monitoring time to 5, and turned off options like shader cache and surface optimization.

I can now run a prp test of 216091 and 756839. This is progress. However, running a PRP with a 5M FFT fails.

Looking for more Windows "features" I can turn off....

[kriesel]

Originally Posted by Prime95

I ran a two hour mfakto run without problem.

I tried some suggestions from here, namely eliminating game bar. In Wattman I increased the performance monitoring time to 5, and turned off options like shader cache and surface optimization.

I can now run a prp test of 216091 and 756839. This is progress. However, running a PRP with a 5M FFT fails.

Looking for more Windows "features" I can turn off....

I'll see you a shader caching, a surface optimization, and 5 seconds of performance monitoring time, and raise you another 5 seconds of performance monitoring time, and a chill, a -carry long option, +32GB of ram, and a BIOS change ("balanced performance" to "performance")

With no overclocking attempted:
V6.5-84-g30c0508, which is type 1 residues, but no P-1 on NVIDIA
gpuowl-win -user kriesel -cpu roa/radeonvii -use FMA_X2 -device 1 -carry long
or ORIG_X2, speeds are about the same, and I was able to successfully run brief trials at a subset of known mersenne primes selected for about 2:1 exponent spacing; others not attempted.

For PRP=0,1,2,82589933,-1,76,0,3,1
GECs pass,
and the 1M iteration res64 matched the respective entry at https://www.mersenneforum.org/showpo...83&postcount=5

Trying v6.11-9 after that, it matched the 10,000,000 iteration res64 from a separate start for 82589933.
Radeon VII fan speed is humming along at 61%, one of the few parameters gpu-z will show in remote desktop.

[kriesel]

Originally Posted by kriesel

Trying v6.11-9 after that, it matched the 10,000,000 iteration res64 from a separate start for 82589933.
Radeon VII fan speed is humming along at 61%, one of the few parameters gpu-z will show in remote desktop.

I found it hung this morning, at 34.8% complete, only lost a couple of hours. No GEC errors.
Required a system restart to recover.
The good news is I bumped the memory clock up 2% and turned off the annoying game slideshow in Wattman, and gpuowl v0.5 (the only gpu LL tester with the Jacobi check included) completed all the packaged selftests of v0.5 gpuowl with no detected error. That would make some fast LL DC results, at ~0.86 ms/iteration single instance.

This v0.5 selftest sequence slams the gpu power pretty hard, fluctuating it between 18. W min and 258. W max as reported by GPU-Z between and during brief selftests.

[Prime95]

My linux setups are all configured for best throughput at reasonable power consumption. That translates to about 1550MHz core, 870mV and 1150MHz memory. Power draw ~180 watts. YMMV.

Best throughput comes from running 2 PRP tests at the same time.

[kriesel]

Gpuowl v0.5 LL DC on a 50M exponent and gpuowl v6.11-9 on PRP Mp51* sharing the same XFX Radeon VII, at ~1700Mhz auto set gpu clock, 1075Mhz memory clock, stock voltage, ok, and averaging 187W. Changed the V6.11-9 PRP run from -carry long to -carry short, and seemingly instantly got a display hang.

Patient: It hurts when I do this.
Doctor: Don't do that.

[ewmayer]

Originally Posted by kriesel

Patient: It hurts when I do this.
Doctor: Don't do that.

You left off the ending of the story: "Patient stops breathing, expires a few minutes later."

[kriesel]

Originally Posted by ewmayer

You left off the ending of the story: "Patient stops breathing, expires a few minutes later."

This one stopped breathing, but was easily and quickly revived.