4 ms·

Not just silicon miracle. Especially a driver miracle. Right now AMD is kinda usable-ish on DX and so and so in OpenCL if one spends a lot of time optimizing s

Not just silicon miracle. Especially a driver miracle. Right now AMD is kinda usable-ish on DX and so and so in OpenCL if one spends a lot of time optimizing specifically for them. Their OpenGL support is atrocious. Their Vulkan implementation is full of bugs. AMD has had issues with drivers for ages. And there is no improvement in sight.

avinium 7y ago
I get the feeling that OpenCL is going to become a second-class citizen in favour of ROCm (https://rocm.github.io/ https://rocm.github.io/), which looks very promising for HPC/deep learning on AMD GPUs. I haven't got stuck into it yet, but when I was doing some research it seemed like the AMD team had invested a lot of effort in minimizing the work needed to either support/port CUDA-specific code. I'm optimistic about AMD fighting their way back into contention as far as deep learning goes. I don't know what that means in the gaming world, but I assume a win in the former will help them in the latter.
- antt 7y ago
  Do you have an link that goes over what ROCm does and why I should care about it? I'm interested in getting a purely amd setup this year.
  - avinium 7y ago
    Here's a link that's a couple of years old now, but gives a pretty reasonable overview of where it sits in the stack: https://gpuopen.com/ported-caffe-hip-heres-happened/ https://gpuopen.com/ported-caffe-hip-heres-happened/ As to why you should care, it depends at what level you work. I'm not a graphics/HPC programmer, so I can't really comment on what you'd be using it for in those areas. For DL applications, though, CUDA/CUDNN used to be the only GPU libraries that offered some of the common/optimized operations you'd encounter in deep learning (e.g. 2d convolutions). That meant you were limited to NVIDIA GPUs within all the higher level frameworks (Tensorflow/Pytorch/etc). With ROCm more developed, it can supplant CUDA as the low-level library for those frameworks. This means you can start using AMD GPUs for training/inference, which are considerably cheaper. In fact, in the few months that have passed since I last checked, it now appears that ROCm supports TF - big step. it means you're no longer locked to NVIDIA GPUs
- anfilt 7y ago
  I would prefer anything that takes off that is not the proprietary lock that is CUDA.

avinium 7y ago

I get the feeling that OpenCL is going to become a second-class citizen in favour of ROCm (https://rocm.github.io/ https://rocm.github.io/), which looks very promising for HPC/deep learning on AMD GPUs. I haven't got stuck into it yet, but when I was doing some research it seemed like the AMD team had invested a lot of effort in minimizing the work needed to either support/port CUDA-specific code. I'm optimistic about AMD fighting their way back into contention as far as deep learning goes. I don't know what that means in the gaming world, but I assume a win in the former will help them in the latter.

antt 7y ago
Do you have an link that goes over what ROCm does and why I should care about it? I'm interested in getting a purely amd setup this year.
- avinium 7y ago
  Here's a link that's a couple of years old now, but gives a pretty reasonable overview of where it sits in the stack: https://gpuopen.com/ported-caffe-hip-heres-happened/ https://gpuopen.com/ported-caffe-hip-heres-happened/ As to why you should care, it depends at what level you work. I'm not a graphics/HPC programmer, so I can't really comment on what you'd be using it for in those areas. For DL applications, though, CUDA/CUDNN used to be the only GPU libraries that offered some of the common/optimized operations you'd encounter in deep learning (e.g. 2d convolutions). That meant you were limited to NVIDIA GPUs within all the higher level frameworks (Tensorflow/Pytorch/etc). With ROCm more developed, it can supplant CUDA as the low-level library for those frameworks. This means you can start using AMD GPUs for training/inference, which are considerably cheaper. In fact, in the few months that have passed since I last checked, it now appears that ROCm supports TF - big step. it means you're no longer locked to NVIDIA GPUs
anfilt 7y ago
I would prefer anything that takes off that is not the proprietary lock that is CUDA.