NVIDIA vs. Radeon Linux 5.0 + Mesa 19.0 Drivers - 14-Way Gaming GPU Comparison

[shmerl]

RadeonSI has always been a little faster than Vulkan, but Michael doesn't test it as much as Vulkan. I'm always looking forward to some Mad Max OpenGL benchmarks of RadeonSI vs NVIDIA, where we know that RadeonSI is much faster than RADV.

We also need more Wine / dxvk benchmarks, that's the most common Vulkan use cases these days.

[dungeon]

Per watt and per dollar metrics will be in the VII review.

VII should be in between RTX 2070 and 2080, while in couple years thanks to the unofficial AMD Fine Wine tech that might be even somewhere between RTX 2080 and 2080 Ti

But, lets wait that 7 February... things might not even run

[puleglot]

AMD Fine Wine tech

That's why they are red.

[dungeon]

Is it just me, or has the red team seen some serious performance increases in several benchmarks? Where did that come from all of a sudden? LLVM9?

As i see Padoka PPA uses Autotools, while Oibaf is on Meson build... second have assertions enabled and this one disabled... among other things, didn't look further as with different builds perf diff could be much unrelated to the actual driver changes

[M@GOid]

RadeonSI has always been a little faster than Vulkan, but Michael doesn't test it as much as Vulkan. I'm always looking forward to some Mad Max OpenGL benchmarks of RadeonSI vs NVIDIA, where we know that RadeonSI is much faster than RADV.

You heard the man Michael.

Make it happen.

[BNieuwenhuizen]

RadeonSI has always been a little faster than Vulkan, but Michael doesn't test it as much as Vulkan. I'm always looking forward to some Mad Max OpenGL benchmarks of RadeonSI vs NVIDIA, where we know that RadeonSI is much faster than RADV.

I don't know, I think it is pretty close: https://openbenchmarking.org/result/...AL-MADMAX20179

[TheYoshiGuy]

These Vega performance improvements are all well and good, but they're still a totally broken, hopeless set of drivers if you're trying to run any of the popular emulators.
1. If you use the Wii U emulator (Cemu) to try and run Mario Kart 8, the entire system freezes. The bug report is a year old, no solutions: https://bugs.freedesktop.org/show_bug.cgi?id=105251 Some of us wasted a lot of time submitting logs / compiling patched kernels/drivers here.
2. If you try to run Super Mario Odysee and other games in the Nintendo Switch emulator (Yuzu) with a Vega or Polaris card, the entire system locks up.
3. If you try to even start the other Wii U emulator (decaf) with any AMD card, it crashes immediately.
4. This is also Vega-only, if you try to run Breath Of The Wild in the latest version of Cemu, you'll get game-breaking bugs. Polaris is not affected.
5. If you try to render PS3 games in rpcs3 at 4k resolution, you get graphics artifacts (Polaris and Vega). Nvidia cards are fine. This is a RADV issue, AMDVLK is fine though so luckily we have that option.

There's also the general performance issues, where I can't even match the framerate of someone with a GTX 1060 when using any of these OpenGL-based console emulators.

I switch from Nvidia to AMD last year when the Vega64 came out, after reading how great AMD are on Linux now. Worst PC hardware decision I've ever made.

[theriddick]

The Vega64 is doing quite well, shame its not a 250W card tho.
Wish AMD would limit their TDP to 250W, 300W is just silly IMO because you end up with allot of heat and a very large card and heatsink with no small form factor options.

Also I think the RTX2060 performs better under windows last time I checked, it was on the heels of the 2070! Maybe NVIDIA needs to do some work with the linux drivers for that card?

These Vega performance improvements are all well and good, but they're still a totally broken, hopeless set of drivers.

So those specific emulators has issues with the card? and you know for damn sure that VRM/CORE/Memory isn't overheating?

[dungeon]

5. If you try to render PS3 games in rpcs3 at 4k resolution, you get graphics artifacts (Polaris and Vega). Nvidia cards are fine. This is a RADV issue, AMDVLK is fine though so luckily we have that option.

Likely LLVM related bug, since AMDVLK uses something slightly else

[microcode]

Harmonic mean is an interesting choice. :- )