Radeon VII (2nd gen consumer Vega GPU)

[ewmayer]

Originally Posted by xx005fs

Actually not bad when considering brand new, but the pro cards can't be overclocked so it will for sure be memory bound. However, RVII looks to be a much better deal when buying used.

And if the Pro finds a market, we can hope that the older R7s will start coming onto the used-gear market in decent volume, thus pushing prices for those down a bit. Or am I being overoptimistic?

[xx005fs]

Originally Posted by ewmayer

And if the Pro finds a market, we can hope that the older R7s will start coming onto the used-gear market in decent volume, thus pushing prices for those down a bit. Or am I being overoptimistic?

That's certainly possible, but I don't think many workstation users will use Radeon Pros over Quadros. However, I am sure that the release of new gaming cards will make the Radeon VII gamers dump them for the new rx 6000 or rtx 3000 series



On a side note, with Nvidia's announcement of Ampere GA100 today, if they by chance release a card similar to the Titan V called the Titan A or something that has all of the capability of the server GA100 (maybe with 1 less stack of HBM2e), then it will be an absolute beast in PRP, but I certainly don't expect the price to drop.

[Viliam Furik]

My settings are:
Voltage: 975 mV
Power limit: -15%
Core clock: 1450 MHz
Memory: 1100 MHz

I am applying them with the MSI Afterburner software.

Temperatures are:
GPU temp: about 75 °C
Hot spot temp: about 95 °C

As I mentioned in the DC and TC thread, I have done two LLDC tests, neither of them matched. TC is in progress. I have done third LLDC, on a known residue (50000059), this time it matched.

Can somebody give me advice on how to change the underclock to not give errors, and maybe also how to keep it cooler at the same time?

BTW both not-matched tests had one Jacobi error in them. The matching one also had one Jacobi error, but that was obviously fixed. I can now only hope that this was the same scenario as in the first two tests.

I will give a status update when the TCs are finished.

[preda]

Originally Posted by Viliam Furik

My settings are:
Voltage: 975 mV
Power limit: -15%
Core clock: 1450 MHz
Memory: 1100 MHz

I am applying them with the MSI Afterburner software.

Temperatures are:
GPU temp: about 75 °C
Hot spot temp: about 95 °C

As I mentioned in the DC and TC thread, I have done two LLDC tests, neither of them matched. TC is in progress. I have done third LLDC, on a known residue (50000059), this time it matched.

Can somebody give me advice on how to change the underclock to not give errors, and maybe also how to keep it cooler at the same time?

BTW both not-matched tests had one Jacobi error in them. The matching one also had one Jacobi error, but that was obviously fixed. I can now only hope that this was the same scenario as in the first two tests.

I will give a status update when the TCs are finished.

Please do PRP testing while you tune the GPU setup. When it's rock-solid (less than one PRP error per week) you can switch to LL.

It's possible that the MEM overclock is too high. Or that the voltage too low. To keep it cool, lower the sclk (core clock).

[ewmayer]

Viliam, your core-clock of 1450MHz is roughly midway between what an R7 runs at under Linux/ROCm at sclk =3 (1373MHz) and 4 (1547MHz). I've had no problems running any of my 4 R7s at those clocks and mclk = 1150MHz, but I could see one with a slightly-lesser-grade chunk of Silicon causing problems at your settings. Echoing Mihai, I suggest:

1. Reduce your core-clock to, say, 1350MHz;
2. What is your system-default mclk frequency? I'm guessing it's < 1100MHz, suggest you revert to it and...
3. Do several PRPs - preferably PRP-DCs so you can be 100% confident in an "error free" run really having been so - and use the much-better error-checking of those to tune your hardware.

[Prime95]

Originally Posted by Viliam Furik

Can somebody give me advice on how to change the underclock to not give errors, and maybe also how to keep it cooler at the same time?

Not specific advice to your question:

Consider a Linux setup. AMD's linux driver and opencl compiler are *way* better (faster, more reliable) in the Linux environment.

[Prime95]

For sale (not from me): A used Radeon VII, $550

https://hardforum.com/threads/fs-rad...-16gb.1998015/

[kriesel]

Thanks for taking reliability seriously. (Some of the following may be quite obvious to you. If so, consider it might not be to someone after you, and help them.)

PRP with the highly reliable GEC to detect errors is the way to go to find settings for which your gpu is stable. Establish stability at stock memory clock rate first. If you don't have that, you probably won't at anything faster. If not too much time has passed, and it's not stable at 1000Mhz, you should be able to exchange it for one that is, or get a refund and buy one from somewhere else.
Reducing gpu clock rate reduces power and indirectly reduces temperature. Cooler electronics live longer and can be more reliable.
Lifting the high end of the fan curve may help temperatures with less impact on throughput.
On Windows the driver install package often has a utility for doing all that. Save the tuned profile of settings for reuse. On linux Ernst's thread on setting up gpuowl for linux in the gpuowl subforum may help.

After you find and save and set at startup, settings that provide reliability, then you may be able to run LL or P-1 reliably, and verify that by doing some LL DC that matches and P-1 on test exponents with known factors.
Electronics change with age. I suggest retesting for reliability and reevaluating annually.

Good luck.

[kriesel]

Expected availability end of August. https://www.bhphotovideo.com/c/produ...vii_16gb.html/
Note "noncancelable and nonreturnable".

[ewmayer]

Originally Posted by kriesel

Expected availability end of August. https://www.bhphotovideo.com/c/produ...vii_16gb.html/
Note "noncancelable and nonreturnable".

End of august, $1899, with those terms? No thanks. Amazon just undercut that by $400, free-ship, "Arrives: Aug 10 - 12; Fastest delivery: Aug 7 - 11. Only 9 left in stock - order soon."

They also have this, which I'd never seen before:

AMD Radeon Vega Frontier Edition Liquid Retail

Is this basically an R7 w/liquid-cooling? Because it's priced $890 new, ~$800 from the cheapest reseller, both lower than used R7s have been going for of late. Above is w/free-shipping, and free returns.

[Runtime Error]

Originally Posted by ewmayer

End of august, $1899, with those terms? No thanks. Amazon just undercut that by $400, free-ship, "Arrives: Aug 10 - 12; Fastest delivery: Aug 7 - 11. Only 9 left in stock - order soon."

That's for the Radeon Pro WX 9100, not the Radeon Pro VII. Maybe you copy/pasted a different link?

Pro WX 9100: https://www.amd.com/en/products/prof...-liquid-cooled
Pro VII: https://www.amd.com/en/products/prof...radeon-pro-vii

Originally Posted by ewmayer

They also have this, which I'd never seen before:

AMD Radeon Vega Frontier Edition Liquid Retail

Is this basically an R7 w/liquid-cooling? Because it's priced $890 new, ~$800 from the cheapest reseller, both lower than used R7s have been going for of late. Above is w/free-shipping, and free returns.

Not an R7. The specs say it has 819 GFLOPs double precision, while R7s have 3.46 TFLOPs double precision:

VF: https://www.amd.com/en/products/prof...ion-air-cooled
R7: https://www.amd.com/en/products/graphics/amd-radeon-vii

Edit: Scroll to the bottom of these pages for specs.