NVIDIA vs. Radeon Vulkan & OpenGL Performance With A Celeron, Pentium & Core i7

[phoronix]

Phoronix: NVIDIA vs. Radeon Vulkan & OpenGL Performance With A Celeron, Pentium & Core i7

Here is an interesting OpenGL vs. Vulkan Linux benchmark comparison where I take two competing NVIDIA and AMD cards, the Radeon RX 580 and GeForce GTX 1060, and test the available benchmark-friendly OpenGL/Vulkan Linux games while doing these tests each on an Intel Celeron, Pentium, and Core i7 processors in looking at the performance scaling.

http://www.phoronix.com/vr.php?view=24932

[tildearrow]

Typo:

Dawn of War 3 with loe quality settings

[chuckula]

Interesting to see that Vulkan's biggest positive impact is for lower core count and lower performance CPUs that had more issues with OpenGL.

Not a particular surprise when you read the rationale behind why the newer APIs were put together in the first place, but I think there was a big misconception that Vulkan meant you would be running around gaming with a 16 core server CPU all of the sudden when in fact the real benefits are for lower end dual-core parts.

[bridgman]

... when in fact the real benefits are for lower end dual-core parts.

Can I tweak that slightly ? The real benefits are for lower end CPUs relative to GPU performance. In the Intel world that means lower end dual-core parts with high-ish clocks (or also our Stoney Ridge APU) but it could just as easily mean more even lower clocked cores eg. current ARM parts.

[jrch2k8]

Interesting to see that Vulkan's biggest positive impact is for lower core count and lower performance CPUs that had more issues with OpenGL.

Not a particular surprise when you read the rationale behind why the newer APIs were put together in the first place, but I think there was a big misconception that Vulkan meant you would be running around gaming with a 16 core server CPU all of the sudden when in fact the real benefits are for lower end dual-core parts.

mmm in fact both scenarios are correct, if you have a very weak CPU Vulkan will help simply because its more direct and have less overhead than its counterparts(DX11- and OGL) and is visible even with partial ports(GL + vulkan or simple Vulkan migrations.)

Now when you have extremely strong CPUs 8-128 cores and huge bandwidths and several GPUs(or some uber behemoth GPU running at 6Ghz that can keep up with 128 threads uploading data) Vulkan should allow you extract the last bit of performance out of it but the issue now is moved to the engine and for now most engines only use Vulkan partially(as a drop in replacement, so it mostly work like GL or DX do and very thread limited) because if their engines were developed entirely for 4 threads(including AI, Physics,etc) and DX is nearly impossible to upscale to behemoth mode and keep it compatible with DX11-OGL(you can thank Intel for a decade of milking quad cores) and sadly 99% of game engine are in this position.

but yeah, long story short Vulkan can theoretically scale to something like this:

- 128 threads(EPYC gaming yeah, <-- pun intended): 20 threads handling texture flow, 40 thread handling rendering, 20 threads handling compositing, 40 threads handling compute task, 8 handling other stuff

- 64+ Gb Octa Channel high speed DDR4 + HBM2 on GPUs: ultra wide compute, geometry and texture lazy ring buffers for 4K+ gaming to ultra maximize data residence but with enough bandwidth to keep the beasts feed

- 4+ Vega/Pascal GPU + HBM2: 1 GPU entirely to manage Geometry(in current engines geometry is hard limited tho), 1 GPU to handle physics and other compute intensive task, 2+ GPU to render and composite(Note that vulkan don't handle multi gpu per se as crossfire or SLI(AFR, etc.) but with composite and partial rendering and other techniques is pretty much doable but more complex but at this scale who cares)

The question is when games studio will decide to let DX11-OGL die and focus only on Vulkan-DX12+ with ultra scale

[mphuZ]

Interestingly, will the optimization of OpenGL transfer to Windows ?
On Linux it's almost on par with NVIDIA!

RADV without tears impossible to watch.. Is not enough Pro drivers..

[eydee]

It's incredible how well these mining cards can run games.

[chimpy]

Great tests Michael. It's great to see lower/mainstream consumer CPUs being benchmarked for gaming since that's what I usually buy 😊

[leipero]

I don't know, Celeron G3930 still have 30% single-thread advantage over FX 4300 for example, trowing in some FX CPU would test Vulkan advantage better in my opinion. When radv becomes better ofc, G3930 is simply slow CPU, but using fast FX 6+ CPU with bad IPC should show most advantage in Vulkan when drivers work properly, at least in theory.

[Eliasvan]

Very interesting tests Michael, great job!

It would've also been interesting to include the corresponding AMDGPU-PRO results.