Re:Tensorflow 2.0.0-Alpha0
by dragontamer on Wednesday March 06, 2019 @02:26PM
(#19321808)
ROCm, or more specifically HCC, is comparable to CUDA: AMD needs to develop it more before anything else. Its the fundamental structure to GPU-coding.
AMD can't even write a comparable library unless it is built on top of ROCm / HCC. Like, how would you even start to implement Tensorflow 2.0.0 for AMD Graphics cards otherwise?
OpenCL was theoretically an option. But the old AMDGPU-pro driver had the OpenCL compiler as part of the driver stack. Which means OpenCL Compiler issues would appear and disappear as you updated your graphics card. Uhhhh... not good for deployment to say the least. Lots of OpenCL code with "This causes the compiler to enter an infinite loop on AMD Driver 12.5.whatever".
There are serious issues about OpenCL's fundamental design. You really shouldn't have a full OpenCL compiler hidden inside of your device drivers. Its a nightmare to test and deploy, especially in a world of changing device drivers.
--------
ROCm is a proper compiler stack and a proper development environment. If anything, AMD's mistake was taking so long before ROCm became usable. Even today, there are some issues (it doesn't work with Blender 2.79 OpenCL code yet), but at least its a proper development environment where you can build libraries on top of.
AMD needs to get its compute stack in order if it wants to be taken seriously. Fortunately, it seems like they understand the issue. Hopefully the next year or two will be helpful, since they're ramping up on ROCm development.
AMD can't even write a comparable library unless it is built on top of ROCm / HCC. Like, how would you even start to implement Tensorflow 2.0.0 for AMD Graphics cards otherwise?
OpenCL was theoretically an option. But the old AMDGPU-pro driver had the OpenCL compiler as part of the driver stack. Which means OpenCL Compiler issues would appear and disappear as you updated your graphics card. Uhhhh... not good for deployment to say the least. Lots of OpenCL code with "This causes the compiler to enter an infinite loop on AMD Driver 12.5.whatever".
There are serious issues about OpenCL's fundamental design. You really shouldn't have a full OpenCL compiler hidden inside of your device drivers. Its a nightmare to test and deploy, especially in a world of changing device drivers.
--------
ROCm is a proper compiler stack and a proper development environment. If anything, AMD's mistake was taking so long before ROCm became usable. Even today, there are some issues (it doesn't work with Blender 2.79 OpenCL code yet), but at least its a proper development environment where you can build libraries on top of.
AMD needs to get its compute stack in order if it wants to be taken seriously. Fortunately, it seems like they understand the issue. Hopefully the next year or two will be helpful, since they're ramping up on ROCm development.