3 ms·

AMD really dropped the ball on GPGPU focus, neither for Blender nor any ML framework.....

UK-AL 4y ago
What gets me is that it's really a software thing. You can get rocm working on RDNA 2 consumer cards. It's just a pain, only works linux and when you set it up on linux it's fiddly and can't just use the built in amd stuff in the kernel/mesa.
CodeArtisan 4y ago
AMD GPUs performances on Blender3D are atrocious. 7900xtx (61 tflops) beaten by the 3060ti (16 tflops). AMD apologists have been saying "it will change next version" for years now but nothing has changed so far. https://techgage.com/article/blender-3-4-performance-deep-dive-cycles-eevee-viewport/ https://techgage.com/article/blender-3-4-performance-deep-di...
- salawat 4y ago
  Care to explain how, exactly? Lack of compilable GPGPU kernel primitives?
  - mivance 4y ago
    I would guess shader compiler output microcode quality. It's not hard to imagine that Blender's workloads for "offline" renders are wildly different from game workloads and have not been prioritized in analysis of the shader compiler performance.
- jjoonathan 4y ago
  In 2011, I desperately wanted to buy an AMD card. More perf per dollar, don't feed the nvidia beast, etc etc. Between OpenCL for GPGPU and Blender OpenCL support, I was ready -- only to find that the Blender support didn't work unless you did a custom build and even then it performed so poorly that the CPU would beat it. I sold my AMD card and bought an NVidia one. In 2014, I desperately wanted to buy an AMD card. More perf per dollar, don't feed the nvidia beast, etc etc. AMD had devoted someone to getting the Blender kernels working a few years back, surely now was the time? No, Blender support was still trash. Ok, GPGPU was more important to my use-case anyway. I wasted 2 months trying to write my own compute projects on that AMD GPU. Loads of crashes, leaks (N runs until a black screen and lockup), weird performance issues, debuggers only worked intermittently, it was bad. One time I had spent the whole damn day chasing an OpenCL issue in circles -- only to have the exact, unmodified kernel run flawlessly when I tried it on my buddy's nvidia card. That was it. I sold my AMD card and bought an NVidia one, once more eating the ebay fees, the spread, and the green tax. 10 years later, I hope AMD's success in the CPU space means they now have money and can turn this around. Still, I won't be the guinea pig anymore. I'll buy an AMD card when I see it doing the things I want it to do and not a second earlier. Marketing puffery isn't enough this time. I need to see proof.
- captainbland 4y ago
  I think it's unlikely to change as AMD's strategy is to support GPGPU pretty much only on their CDNA products. Anything Radeon gets is pretty much a happy accident at this point. Unless AMD have a total strategic rethink, Nvidia is going to clean up in GPGPU on the desktop.
- zamadatix 4y ago
  People here are commenting on GPGPU (of which AMD consumer GPUs did split architectures with the compute focused GPUs at the arrival of RDNA) but the performance canyon here is squarely do with OptiX, i.e. using the dedicated ray tracing hardware not the general purpose component. AMD's card has very little dedicated RT hardware while Nvidia's cards have comparatively quite a bit and no realistic amount of general purpose FLOPS is going to be relevant in Blender. You can see evidence of this by looking at the performance in the "No OptiX" chart and seeing the 7900 XTX beats the 3060 Ti there (though maybe not by as much as one might expect).
ur-whale 4y ago
> AMD really dropped the ball on GPGPU focus, neither for Blender nor any ML framework..... They certainly did, and what a tragic strategic mistake that was. AMG GPU's are only good for gaming and nothing else (bad for 3D modeling, bad for GPGPU, bad for studio type work, etc...) From a pure marketing standpoint, this makes them look like cheap plastic toys, whereas NVidia GPU can not only do gaming superbly, they can also do GPGPU which brings the whole of scientific computing and ML into play and gives them a sheen of seriousness and solidity that AMD simply can't match. Perception may not be everything, but it certainly is a huge part of everything.
rbanffy 4y ago
What is missing from it? It doesn't seem to be raw performance, at least.
- shmerl 4y ago
  Vulkan support in Blender is missing.
captainbland 4y ago
It's less that they dropped the ball so much as locked it behind a large gate embossed with a dollar symbol and behind it live compute monsters like the Instinct MI2XX line/CDNA2 (and soon their successors!). But in the consumer space, yeah, Nvidia is a no-brainer if you want GPGPU.
- mardifoufs 4y ago
  But trying to use MI2XX compute is so hard even on public clouds. I have tried it for a bit and it always seems to throw some kind of error if you want to run anything more "generic" on it (like say, pytorch) , as opposed to more custom, finely tuned code you'd find on usual HPC.
  - slavik81 4y ago
    I didn't know there was a public cloud with MI200 instances available. Where is that?
    - mardifoufs 4y ago
      Oops you're right, it was MI25 instances on azure. Though oddly enough azure has announced Mi200 instances almost a year ago but they still aren't accessible.

What gets me is that it's really a software thing. You can get rocm working on RDNA 2 consumer cards. It's just a pain, only works linux and when you set it up on linux it's fiddly and can't just use the built in amd stuff in the kernel/mesa.

CodeArtisan 4y ago

AMD GPUs performances on Blender3D are atrocious. 7900xtx (61 tflops) beaten by the 3060ti (16 tflops). AMD apologists have been saying "it will change next version" for years now but nothing has changed so far. https://techgage.com/article/blender-3-4-performance-deep-dive-cycles-eevee-viewport/ https://techgage.com/article/blender-3-4-performance-deep-di...

salawat 4y ago
Care to explain how, exactly? Lack of compilable GPGPU kernel primitives?
- mivance 4y ago
  I would guess shader compiler output microcode quality. It's not hard to imagine that Blender's workloads for "offline" renders are wildly different from game workloads and have not been prioritized in analysis of the shader compiler performance.
jjoonathan 4y ago
In 2011, I desperately wanted to buy an AMD card. More perf per dollar, don't feed the nvidia beast, etc etc. Between OpenCL for GPGPU and Blender OpenCL support, I was ready -- only to find that the Blender support didn't work unless you did a custom build and even then it performed so poorly that the CPU would beat it. I sold my AMD card and bought an NVidia one. In 2014, I desperately wanted to buy an AMD card. More perf per dollar, don't feed the nvidia beast, etc etc. AMD had devoted someone to getting the Blender kernels working a few years back, surely now was the time? No, Blender support was still trash. Ok, GPGPU was more important to my use-case anyway. I wasted 2 months trying to write my own compute projects on that AMD GPU. Loads of crashes, leaks (N runs until a black screen and lockup), weird performance issues, debuggers only worked intermittently, it was bad. One time I had spent the whole damn day chasing an OpenCL issue in circles -- only to have the exact, unmodified kernel run flawlessly when I tried it on my buddy's nvidia card. That was it. I sold my AMD card and bought an NVidia one, once more eating the ebay fees, the spread, and the green tax. 10 years later, I hope AMD's success in the CPU space means they now have money and can turn this around. Still, I won't be the guinea pig anymore. I'll buy an AMD card when I see it doing the things I want it to do and not a second earlier. Marketing puffery isn't enough this time. I need to see proof.
captainbland 4y ago
I think it's unlikely to change as AMD's strategy is to support GPGPU pretty much only on their CDNA products. Anything Radeon gets is pretty much a happy accident at this point. Unless AMD have a total strategic rethink, Nvidia is going to clean up in GPGPU on the desktop.
zamadatix 4y ago
People here are commenting on GPGPU (of which AMD consumer GPUs did split architectures with the compute focused GPUs at the arrival of RDNA) but the performance canyon here is squarely do with OptiX, i.e. using the dedicated ray tracing hardware not the general purpose component. AMD's card has very little dedicated RT hardware while Nvidia's cards have comparatively quite a bit and no realistic amount of general purpose FLOPS is going to be relevant in Blender. You can see evidence of this by looking at the performance in the "No OptiX" chart and seeing the 7900 XTX beats the 3060 Ti there (though maybe not by as much as one might expect).

ur-whale 4y ago

> AMD really dropped the ball on GPGPU focus, neither for Blender nor any ML framework..... They certainly did, and what a tragic strategic mistake that was. AMG GPU's are only good for gaming and nothing else (bad for 3D modeling, bad for GPGPU, bad for studio type work, etc...) From a pure marketing standpoint, this makes them look like cheap plastic toys, whereas NVidia GPU can not only do gaming superbly, they can also do GPGPU which brings the whole of scientific computing and ML into play and gives them a sheen of seriousness and solidity that AMD simply can't match. Perception may not be everything, but it certainly is a huge part of everything.

rbanffy 4y ago

What is missing from it? It doesn't seem to be raw performance, at least.

shmerl 4y ago
Vulkan support in Blender is missing.

captainbland 4y ago

It's less that they dropped the ball so much as locked it behind a large gate embossed with a dollar symbol and behind it live compute monsters like the Instinct MI2XX line/CDNA2 (and soon their successors!). But in the consumer space, yeah, Nvidia is a no-brainer if you want GPGPU.

mardifoufs 4y ago
But trying to use MI2XX compute is so hard even on public clouds. I have tried it for a bit and it always seems to throw some kind of error if you want to run anything more "generic" on it (like say, pytorch) , as opposed to more custom, finely tuned code you'd find on usual HPC.
- slavik81 4y ago
  I didn't know there was a public cloud with MI200 instances available. Where is that?
  - mardifoufs 4y ago
    Oops you're right, it was MI25 instances on azure. Though oddly enough azure has announced Mi200 instances almost a year ago but they still aren't accessible.