Radeon RX Vega On Linux: High-Performance GPUs & Open-Source No Longer An Oxymoron

[AnassAhmed]

Let's save up for the next Vega RX 64.

[juno]

This is such a shame really, even if limited to just 2 instances, it would have been amazing.

Really a shame, even Intel has better virtualization features for their GPU hardware.
Also, I think the Vega cards are expensive enough for their actual performance. And there is still the "Pro" killer feature 16 GiB of memory.

[worohregger]

AFAIK the problem is that it is all or nothing, no ability to say "only enable a couple of VF's".

I have a couple of requests in to see if a cut-down version is possible though.

Thank you, this would be great :-)

[Karbowiak]

AFAIK the problem is that it is all or nothing, no ability to say "only enable a couple of VF's".

I have a couple of requests in to see if a cut-down version is possible though.

I will donate a case of beer if you're successful

Surely others will join in on that

[chithanh]

"only enable a couple of VF's".

From what I have read, support for a single VM instance would already cover the main use cases from consumers and developers:

1. occasionally fire up a Windows VM for gaming
2. developing a cross-platform application and checking if it performs as intended on another platform, without having to reboot

Especially the second one left buyers of Vega FE scratching their heads - why would AMD intentionally disable such a useful feature for development on a card which is specifically marketed at developers?

[Panek]

bridgman Thank you for checking the possibility of a cut-down version. I'd love this to have a Linux host + 1 Windows VM for gaming without the need for 2 GPUs. It sure would make me buy a Vega 56. There's a huge interest in this feature all over the AMD Subreddit, too (e.g. here https://www.reddit.com/r/Amd/comment...e_vega6456_an/ ). Please keep us updated what you'll find.

[Shevchen]

I bought an RX Vega64 and are fine with its performance under Windows 10. After 2 days of tweaking, testing, validating, re-testing and re-tweaking, I finally found its most performant long term sweet spot, which is:
+ 30% Power Target, 1672 MHz core (1.125V), 1060 MHz HBM2 (Voltage control doesn't seem to work with the current driver) and the Firestrike results on Windows are en par with an overclocked GTX 1080. (24650 graphics score)

I could push the card even higher (1752 MHz and 1100 HBM2), but the HBM2 is temp sensitive and the power leakage at 70°C starts to make the card more inefficient, so I had to drop the clock a little to stabilize it for higher temps. But all in all, still a very good score.

BUT: I still don't have access to the Power-states 1-5. THIS IS A MUST if you want to tweak the cards performance, esp. as its power/temp limited. The cards stock values are nothing I would see as competitive - its heavily nerfing itself with the overvoltage. (and a lot of people who try to tweak it set it to 50% Power Target without knowing, that this is too much for the cooler and wonder, why the card throttles)

Also, as I plan to put this card on water, it already losts its warranty (but I'm pleased with my sample so I just did it). The thermal paste used on the card was like glue. Already hard after the first burn-in test and it didn't seemed like a very good one too. I've replaced it with my Gelid extreme thermal paste and had 5-6°C less.

Now, of course with Linux I could overwrite all the values, but in order to test them, I must stress test it and overclocking/undervolting under Linux is just apita compared to Windows. (Thats why I calibrate it under Windows and then take those calibrates values over to Linux - at least, that was the plan)

So - here are the things I'd say AMD should do on priority:
- Get your Wattman drivers stable and reliable (I found a work-around, that re-applying clocks after HMB2 adaptation does the trick)
- Open up the custom Bios for modding all the P-states - some cards have a massive headroom for undervolting. You are losing customers already because of that