Radeon VII (2nd gen consumer Vega GPU)

[tServo]

Originally Posted by dcheuk

Well thanks for the suggestion and help. Tomorrow I'll poke around and if nothing works then will pull one card out as M344587487 suggested and see what happens.

Before you do that, you should:

(1) Use George's version found in post:
https://www.mersenneforum.org/showth...win#post530859

and
(2) run it with -use ORIG_X2 on the command line.

You may also want to use -yield to reduce cpu usage. It runs slightly slower, however.

I have almost an exact setup as you and this works for me.

[tServo]

One more thing dcheuk,

I forgot to mention that you need a fairly new Nvidia driver that has support for OpenCl 2.0 .

If it's numbered above 400.xx, it should be OK but you might want to get the latest one anyway.

I believe I have great luck with one numbered 441.xx

[dcheuk]

Originally Posted by tServo

Before you do that, you should:

(1) Use George's version found in post:
https://www.mersenneforum.org/showth...win#post530859

and
(2) run it with -use ORIG_X2 on the command line.

You may also want to use -yield to reduce cpu usage. It runs slightly slower, however.

I have almost an exact setup as you and this works for me.

Thanks for everyone's help and suggestions. I found and solved the problem!

This is going to sound silly. The problem originates from the outdated opencl.dll that I put into the directory. After removing that file, everything runs correctly.


Now, I noticed that these gpus gets super hot.

Originally at stock setting, temperature shoots up to 110C (!!!):

2019-12-05 14:49:15 OpenCL compilation in 1609 ms
2019-12-05 14:49:18 94609943 OK     1000   0.00%;  886 us/sq; ETA 0d 23:17; 58bafcc7840f380e (check 0.61s)
2019-12-05 14:49:26 94609943       10000   0.01%;  886 us/sq; ETA 0d 23:16; c5cdcefb7a0c847b
2019-12-05 14:49:35 94609943       20000   0.02%;  881 us/sq; ETA 0d 23:09; f9223716e06e4e56
2019-12-05 14:49:43 94609943       30000   0.03%;  881 us/sq; ETA 0d 23:08; c5fc2f4747198494
2019-12-05 14:49:52 94609943       40000   0.04%;  880 us/sq; ETA 0d 23:07; ca1c893cb4c75e3c
2019-12-05 14:50:01 94609943       50000   0.05%;  880 us/sq; ETA 0d 23:07; 5d63d3b21dc37784
2019-12-05 14:50:10 94609943       60000   0.06%;  880 us/sq; ETA 0d 23:07; 52778c4d5657d79c
2019-12-05 14:50:19 94609943       70000   0.07%;  881 us/sq; ETA 0d 23:08; 700e30b14e5a4d15
2019-12-05 14:50:22 Stopping, please wait..
2019-12-05 14:50:22 94609943 OK    73500   0.08%;  881 us/sq; ETA 0d 23:09; 1a88f4b7c5c725b8 (check 0.59s)

Now undervolt to 1460mhz:

2019-12-05 15:02:17 94609943      800000   0.85%;  965 us/sq; ETA 1d 01:09; ceb1808e1032b07d
2019-12-05 15:02:27 94609943      810000   0.86%;  953 us/sq; ETA 1d 00:50; 1adccde674bd3e65
2019-12-05 15:02:37 94609943      820000   0.87%;  978 us/sq; ETA 1d 01:29; 0856fd235d3fbcfb
2019-12-05 15:02:46 94609943      830000   0.88%;  966 us/sq; ETA 1d 01:10; 6b4eccd1d859cc44
2019-12-05 15:02:56 94609943      840000   0.89%;  959 us/sq; ETA 1d 00:59; e021b2b2f8019f13
2019-12-05 15:03:05 94609943      850000   0.90%;  948 us/sq; ETA 1d 00:41; cd7ceac866f34f67
2019-12-05 15:03:15 94609943      860000   0.91%;  952 us/sq; ETA 1d 00:47; d84ac98a53e7db37
2019-12-05 15:03:24 94609943      870000   0.92%;  947 us/sq; ETA 1d 00:40; a95368c74f62fac6
2019-12-05 15:03:34 94609943      880000   0.93%;  954 us/sq; ETA 1d 00:51; 61adb093cc2334dd
2019-12-05 15:03:43 94609943      890000   0.94%;  950 us/sq; ETA 1d 00:44; c7711c35ddc2bf1e

[dcheuk]

Originally Posted by tServo

One more thing dcheuk,

I forgot to mention that you need a fairly new Nvidia driver that has support for OpenCl 2.0 .

If it's numbered above 400.xx, it should be OK but you might want to get the latest one anyway.

I believe I have great luck with one numbered 441.xx

That's good to know. Sorry I was literally typing the post up while you posted this. Thanks man, now I am a proud owner of Radeon VII.

P.S. I found it funny that I got two identical cards from the same vendor, and one of them has samsung memory the other hynix. While the samsung gpu, in which I also use for my display, runs gpuowl faster than the idle hynix radeon. (950 us/sq vs 1050 us/sq)

[Prime95]

Originally Posted by dcheuk

Now, I noticed that these gpus gets super hot.

In Wattman, set max speed to 1580, set voltage to 940, set memory clock to 1200.

If errors occur, reduce memory clock. If no errors for a day, try reducing voltage.

[dcheuk]

Originally Posted by Prime95

In Wattman, set max speed to 1580, set voltage to 940, set memory clock to 1200.

If errors occur, reduce memory clock. If no errors for a day, try reducing voltage.

After applying those settings my screen start tearing (see photo and those small pixels) then within seconds it crashed and rebooted hahahaha. Now it's back to original settings.

Both cards are XFX.

Update: Same thing happened at 1175, 1150 and 1100, seems okay at 1050.

[Prime95]

Originally Posted by dcheuk

Same thing happened at 1175, 1150 and 1100, seems okay at 1050.

Ouch. My worst card works at 1150.

[dcheuk]

Got error on 1050mhz and also 1025. Now reverting back to 1000. This is the one with Samsung memory, got the same type of problem twice within 10%. Meanwhile. the one with Hynix was able to do 1100mhz for now without any error, now at 30%.

The errors look like this:

2019-12-05 18:41:01 94609943     6880000   7.27%; 1059 us/sq; ETA 1d 01:48; a24836c6b6fd5231
2019-12-05 18:41:12 94609943     6890000   7.28%; 1059 us/sq; ETA 1d 01:48; 0000000000000000
2019-12-05 18:41:22 94609943     6900000   7.29%; 1057 us/sq; ETA 1d 01:46; 0000000000000000
2019-12-05 18:41:33 94609943     6910000   7.30%; 1057 us/sq; ETA 1d 01:46; 0000000000000000
2019-12-05 18:41:43 94609943     6920000   7.31%; 1057 us/sq; ETA 1d 01:45; 0000000000000000
2019-12-05 18:41:54 94609943     6930000   7.32%; 1058 us/sq; ETA 1d 01:46; 0000000000000000
2019-12-05 18:42:05 94609943     6940000   7.34%; 1057 us/sq; ETA 1d 01:45; 0000000000000000
2019-12-05 18:42:15 94609943     6950000   7.35%; 1057 us/sq; ETA 1d 01:45; 0000000000000000
2019-12-05 18:42:26 94609943     6960000   7.36%; 1057 us/sq; ETA 1d 01:45; 0000000000000000
2019-12-05 18:42:36 94609943     6970000   7.37%; 1057 us/sq; ETA 1d 01:44; 0000000000000000
2019-12-05 18:42:47 94609943     6980000   7.38%; 1057 us/sq; ETA 1d 01:44; 0000000000000000
2019-12-05 18:42:58 94609943     6990000   7.39%; 1057 us/sq; ETA 1d 01:44; 0000000000000000
2019-12-05 18:43:09 94609943 EE  7000000   7.40%; 1057 us/sq; ETA 1d 01:44; 0000000000000000 (check 0.67s)

2019-12-05 20:08:41 89442629     3150000   3.52%; 1069 us/sq; ETA 1d 01:38; 3382781c4c2d8ec3
2019-12-05 20:08:52 89442629     3160000   3.53%; 1067 us/sq; ETA 1d 01:35; 0000000000000000
2019-12-05 20:09:02 89442629     3170000   3.54%; 1068 us/sq; ETA 1d 01:35; 0000000000000000
2019-12-05 20:09:13 89442629     3180000   3.56%; 1067 us/sq; ETA 1d 01:34; 0000000000000000
2019-12-05 20:09:24 89442629     3190000   3.57%; 1067 us/sq; ETA 1d 01:34; 0000000000000000
2019-12-05 20:09:34 89442629     3200000   3.58%; 1068 us/sq; ETA 1d 01:34; 0000000000000000
2019-12-05 20:09:45 89442629     3210000   3.59%; 1067 us/sq; ETA 1d 01:34; 0000000000000000
2019-12-05 20:09:56 89442629     3220000   3.60%; 1067 us/sq; ETA 1d 01:34; 0000000000000000
2019-12-05 20:10:06 89442629     3230000   3.61%; 1067 us/sq; ETA 1d 01:34; 0000000000000000
2019-12-05 20:10:17 89442629     3240000   3.62%; 1067 us/sq; ETA 1d 01:34; 0000000000000000
2019-12-05 20:10:28 89442629 EE  3250000   3.63%; 1067 us/sq; ETA 1d 01:33; 0000000000000000 (check 0.68s)

Would be nice if there's a way to not overwrite the last save file without error, or 4 backup save files.

[LaurV]

Originally Posted by dcheuk

P.S. I found it funny that I got two identical cards from the same vendor, and one of them has samsung memory the other hynix. While the samsung gpu, in which I also use for my display, runs gpuowl faster than the idle hynix radeon. (950 us/sq vs 1050 us/sq)

That may have nothing to do with the memory type (wild ass guess here). Things are similar for all setups I had with Nvidia cards: the one running the monitors is always faster, regardless of SLI, no SLI, socket I put the cards in, card I connect the monitor cables to, or where I put the physics rendering (you can select which card renders the physics), etc, the card connected to the monitors is always a bit faster. Which is bloody counter-intuitive, as that card is quite busy running the screen, it should be slower running TF/LL. The best explanation I could come with was always the fact that the other card goes to some sleep state for some microseconds or so, in certain points between calculations, and it needs time to wake up, while the one driving the monitor has no time to go to sleep, and it is always running full throttle.

[paulunderwood]

Originally Posted by dcheuk

Got error on 1050mhz and also 1025. Now reverting back to 1000. This is the one with Samsung memory, got the same type of problem twice within 10%. Meanwhile. the one with Hynix was able to do 1100mhz for now without any error, now at 30%.

What are your cards temperature readings under load? I have my card set to "setsclk 4" on Linux. Also my case side is removed and the case is on its side to allow the heat to dissipate.

[dcheuk]

Originally Posted by LaurV

the card connected to the monitors is always a bit faster. Which is bloody counter-intuitive

this also holds on my previous system with dual 2080s. The one connected to the monitor is mysteriously faster in both cudalucas and mfaktc.