3 ms·

I think the most interesting bit of this article is this: > So will expanding the use of the Radeon Open Compute (ROCm) development environment, which is the a

I think the most interesting bit of this article is this: > So will expanding the use of the Radeon Open Compute (ROCm) development environment, which is the analog to Nvidia’s CUDA environment. But across the $1.2 billion allocated for the Frontier and El Capitan systems, about $220 million of that is for non-recurring engineering (NRE) costs that will largely be focused on making OpenMP and OpenACC play well with ROCm, which is open source. These HPC centers, and presumably more than a few hyperscalers and cloud builders as well as the Cray unit of HPC that is building the systems and their Linux environment for them, will be kicking into the ROCm software work better and scale further than it does. CUDA may be freely distributed, but it is not open source and neither are Nvidia’s drivers – and in the end, couple to better price/performance for both CPU and GPU compute, the openness of the software may have tipped the balance in AMD’s favor. The GPU vendors were very reluctant to release open documentation for their products for a very long time (and Nvidia still is) because they were afraid it would harm their market position. AMD decided to take that risk (probably helped by already being in a poor market position with Linux deployments) and now they are reaping a reward for their investment in open source software.

dragontamer 7y ago
Indeed. Anyone who has worked with ROCm knows that its... well... rough around the edges. But there's something to be said about "documentation is code", I can pull up Clang and straight up see the parser for ROCm -> AMD GPU compiler, as well as pull up the AMD Vega ISA document (https://developer.amd.com/wp-content/resources/Vega_Shader_ISA_28July2017.pdf https://developer.amd.com/wp-content/resources/Vega_Shader_I...) and learn the assembly language. ROCm is rough around the edges, but it is documented and open source. If all else fails, I can write in assembly and get what I need done. ------- With that being said, NVidia's CUDA is somewhat open too, since their CUDA -> PTX compiler stuff is also in clang, and PTX is also very well documented. I guess people are liking AMD's other open-source parts of ROCm. In particular, ROCm getting upstreamed to the Linux kernel proper for better support. ------- Based on how ROCm has evolved in the past year, they've become more focused. They've cut back on some features (I did like HCC / C++Amp... RIP), and refocused their efforts on what seems to matter. NVidia CUDA continues to march forward of course, but there's clearly a need for a 2nd player, at very least to compete against the sore-spots of NVidia licensing issues and hardware costs.

dragontamer 7y ago

Indeed. Anyone who has worked with ROCm knows that its... well... rough around the edges. But there's something to be said about "documentation is code", I can pull up Clang and straight up see the parser for ROCm -> AMD GPU compiler, as well as pull up the AMD Vega ISA document (https://developer.amd.com/wp-content/resources/Vega_Shader_ISA_28July2017.pdf https://developer.amd.com/wp-content/resources/Vega_Shader_I...) and learn the assembly language. ROCm is rough around the edges, but it is documented and open source. If all else fails, I can write in assembly and get what I need done. ------- With that being said, NVidia's CUDA is somewhat open too, since their CUDA -> PTX compiler stuff is also in clang, and PTX is also very well documented. I guess people are liking AMD's other open-source parts of ROCm. In particular, ROCm getting upstreamed to the Linux kernel proper for better support. ------- Based on how ROCm has evolved in the past year, they've become more focused. They've cut back on some features (I did like HCC / C++Amp... RIP), and refocused their efforts on what seems to matter. NVidia CUDA continues to march forward of course, but there's clearly a need for a 2nd player, at very least to compete against the sore-spots of NVidia licensing issues and hardware costs.