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

[VikingGe]

Vega didn't do a good job at that.

Vega doesn't do a good job at anything except maybe a few compute tasks. As a gaming GPU, it is a complete failure. ~50% more compute power than the old Fiji GPU, barely 30% faster in games. ~100% more raw compute power compared to Polaris, 50% more gaming performance. That's insanely bad scaling, far worse than Fiji, and Fiji was already known to scale quite poorly compared to smaller GPUs.

Earlier Vega benchmarks suggested that this would happen, but it's still disappointing. The only reason Vega might still be an okay-ish option on Linux is that AMD simply has better drivers nowadays. Although it doesn't exactly help that DC isn't merged yet.

[bridgman]

bridgman agd5f, why is the performance of the Vulkan is now very low ?

Most of the performance tuning for Vulkan was done on Windows since we did not yet have large page support in Linux. It took us a while to get that working and we didn't have time to get 2M page support into the AMDGPU-PRO stack before launch.

Once we have it integrated I expect we should see a noticeable performance improvement just from that... and having it will allow the Vulkan team to do more Linux-specific tuning as well.

[Karbowiak]

I've heard that unfortunately it will be available only in pro cards. But may be I missed some update on that.

Noooooooooooooooooo

http://i.imgur.com/HiYlb5N.gifv

[Shevchen]

Vega doesn't do a good job at anything except maybe a few compute tasks. As a gaming GPU, it is a complete failure. ~50% more compute power than the old Fiji GPU, barely 30% faster in games. ~100% more raw compute power compared to Polaris, 50% more gaming performance. That's insanely bad scaling, far worse than Fiji, and Fiji was already known to scale quite poorly compared to smaller GPUs.

Earlier Vega benchmarks suggested that this would happen, but it's still disappointing. The only reason Vega might still be an okay-ish option on Linux is that AMD simply has better drivers nowadays. Although it doesn't exactly help that DC isn't merged yet.

And the reason for the poor scaling is the mostly brute-force approach we render games these days. That we can have the opposite effect shows the first benchmark - and this is even without the ad-hoc improvements we get later on "simply" by switching some features on. Once the Linux driver magic hits the fans and our wizards of code do their chanting, you may be surprised...

But still, I think on both Linux AND Windows, the HBM2 frequency will have a bigger effect on performance than the core frequency.

[mphuZ]

Most of the performance tuning for Vulkan was done on Windows since we did not yet have large page support in Linux. It took us a while to get that working and we didn't have time to get 2M page support into the AMDGPU-PRO stack before launch.

Once we have it integrated I expect we should see a noticeable performance improvement just from that... and having it will allow the Vulkan team to do more Linux-specific tuning as well.

Thank you!
Tell me, please, in the future is it possible to release drivers of the same quality to Windows and Linux? I mean, in one day (Day-1)..

[gerry_here_to_learn]

please test dota2 @ 1080p.

need to know if vega 56/64 can hit 144 fps

[shmerl]

Noooooooooooooooooo
http://i.imgur.com/HiYlb5N.gifv

Yeah, SR-IOV would be neat to have. It will probably allow running Linux guests with accelerated graphics without needing an extra GPU or relying on flaky stuff in Qemu/KVM like virgil and etc.

[R41N3R]

Wow, very promising results! Finally AMD is again leading in some benchmark's, great moment :-) I didn't expect the open source stack is that good at lunch day and this even before a optimization phase for Linux. AMD did work a lot on the driver stack and slowly all the small puzzles will come together.

[bridgman]

Tell me, please, in the future is it possible to release drivers of the same quality to Windows and Linux? I mean, in one day (Day-1)..

Yes, that is what we are aiming for. Vega10 was the first time we tried running Linux drivers on the emulator - there is a fair learning curve and we didn't get very far by the time silicon came back, but for the next generation of chips we are starting on the emulator in parallel with the Windows devs.

[mphuZ]

..but for the next generation of chips we are starting on the emulator in parallel with the Windows devs.

Wow! This is thanks to the open drivers ?
That is, all practices for Windows and Linux are the same ? As I understand it, this will help you to reduce the resources..