4 ms·

This looks lovely and I absolutely love where this is going. However, it would be very cool if deep learning tools weren't dependent on CUDA and could be run o

This looks lovely and I absolutely love where this is going. However, it would be very cool if deep learning tools weren't dependent on CUDA and could be run on, say, any GPU with OpenCL.

blagie 4y ago
It would be lovely if anyone other than NVidia made GPUs useable for GPGPU. I started with Radeon for the open-source factor, but it turned out to be useless in virtually every respect. I bought an NVidia card. I'm now tied to CUDA, but I didn't need to be. I was starting from scratch. The longer before AMD can ship something which actually works, the more entrenched NVidia+CUDA become. At this point, I lost hope.
- atwood22 4y ago
  Would targeting Vulkan instead of CUDA help?
  - naavis 4y ago
    They don't really serve the same purpose. Vulkan is a low-level graphics API while CUDA is a much higher level toolkit for GPGPU.
- blueblob 4y ago
  It's not that they aren't trying, it's that when you reinvent the wheel, you have to do more work. Microsoft is introducing `tensorflow-directml` to avoid this problem by implementing a CUDA equivalent in directX. AMD has ROCm, but it's not well supported because it's not integrated upstream in `tensorflow`.
  - blagie 4y ago
    Well, I bought a ROCm GPU for GPGPU: - I found out I could only use it for compute headless. WTF?!?!?! (https://www.phoronix.com/news/Radeon-ROCm-Non-GUI https://www.phoronix.com/news/Radeon-ROCm-Non-GUI). If it was driving a monitor, my machine would crash hard. There wasn't even an error message. - A lot of other stuff didn't work and just resulted in odd crashes, or worse performance than CPU. I don't know why. - Within 9 months, AMD discontinued support for my card. I raised this as a warranty issue (suitability for advertised purpose), but that obviously would go nowhere without a lawsuit. I had a very expensive brick. - AMD support channels were non-existent. There literally was no way to reach anyone. I bought an NVidia card, and it's been working well ever since. ROCm is not well-supported because it's absolute garbage. You have *less* work reinventing wheels, since it's been invented once. You have more work to get community support and network effects, since you're starting out behind. Fundamentally, though, that can't start to happen if your system doesn't work at all. I agree with you they're trying, but they're trying incompetently. If ROCm was half the speed of CUDA, and wasn't integrating into the latest-greatest frameworks, but it was stable and working, I'd make it work. It wasn't anywhere close to stable and working.
- MonkeyMalarky 4y ago
  More entrenched? The first time I ever touched CUDA was 2012 in university! It's been entrenched for a whole decade!
zepmck 4y ago
You blame CUDA because you have never played AMD + ROCm. Furthermore OpenCL is on dead-end street. With standard programming languages natively supporting CUDA, it has no sense to add extra layers.
jayavanth 4y ago
CPUs are good enough for inference. It will take longer but they work

This looks lovely and I absolutely love where this is going. However, it would be very cool if deep learning tools weren't dependent on CUDA and could be run on, say, any GPU with OpenCL.

blagie 4y ago

It would be lovely if anyone other than NVidia made GPUs useable for GPGPU. I started with Radeon for the open-source factor, but it turned out to be useless in virtually every respect. I bought an NVidia card. I'm now tied to CUDA, but I didn't need to be. I was starting from scratch. The longer before AMD can ship something which actually works, the more entrenched NVidia+CUDA become. At this point, I lost hope.

atwood22 4y ago
Would targeting Vulkan instead of CUDA help?
- naavis 4y ago
  They don't really serve the same purpose. Vulkan is a low-level graphics API while CUDA is a much higher level toolkit for GPGPU.
blueblob 4y ago
It's not that they aren't trying, it's that when you reinvent the wheel, you have to do more work. Microsoft is introducing `tensorflow-directml` to avoid this problem by implementing a CUDA equivalent in directX. AMD has ROCm, but it's not well supported because it's not integrated upstream in `tensorflow`.
- blagie 4y ago
  Well, I bought a ROCm GPU for GPGPU: - I found out I could only use it for compute headless. WTF?!?!?! (https://www.phoronix.com/news/Radeon-ROCm-Non-GUI https://www.phoronix.com/news/Radeon-ROCm-Non-GUI). If it was driving a monitor, my machine would crash hard. There wasn't even an error message. - A lot of other stuff didn't work and just resulted in odd crashes, or worse performance than CPU. I don't know why. - Within 9 months, AMD discontinued support for my card. I raised this as a warranty issue (suitability for advertised purpose), but that obviously would go nowhere without a lawsuit. I had a very expensive brick. - AMD support channels were non-existent. There literally was no way to reach anyone. I bought an NVidia card, and it's been working well ever since. ROCm is not well-supported because it's absolute garbage. You have *less* work reinventing wheels, since it's been invented once. You have more work to get community support and network effects, since you're starting out behind. Fundamentally, though, that can't start to happen if your system doesn't work at all. I agree with you they're trying, but they're trying incompetently. If ROCm was half the speed of CUDA, and wasn't integrating into the latest-greatest frameworks, but it was stable and working, I'd make it work. It wasn't anywhere close to stable and working.
MonkeyMalarky 4y ago
More entrenched? The first time I ever touched CUDA was 2012 in university! It's been entrenched for a whole decade!

zepmck 4y ago

You blame CUDA because you have never played AMD + ROCm. Furthermore OpenCL is on dead-end street. With standard programming languages natively supporting CUDA, it has no sense to add extra layers.

jayavanth 4y ago

CPUs are good enough for inference. It will take longer but they work