3 ms·

This and the fact that AMD now offers tensorflow support for their GPUs: https://www.anandtech.com/show/12032/amd-announces-wider-epyc-availability-and-rocm-17-

This and the fact that AMD now offers tensorflow support for their GPUs: https://www.anandtech.com/show/12032/amd-announces-wider-epyc-availability-and-rocm-17-with-tensorflow-support https://www.anandtech.com/show/12032/amd-announces-wider-epy... Means there is finally some competition in data centers.

dis-sys 9y ago
there is always such competition - E5-2600v4 vs LGA3647 based processors betting on AMD is nothing different from waiting for an open sourced version of Windows from Microsoft.
dragontamer 9y ago
ROCm looks good, but so did OpenCL and I'm not very convinced that OpenCL is a good platform to build code in. ROCm addresses a lot of my concerns on paper, but there isn't much buzz on ROCm... specifically HCC. It looks like ROCm builds off of Microsoft's C++ AMP specification (which also looks like a great idea, but that project unfortunately looks dead). So it provides a "unified" C++ source code that can compile down into x86 OR GPU Assembly (GCN in the case of AMD). Developing OpenCL requires a bit of "copy/paste" going on whenever you start processing data on the GPU vs CPU. Because the OpenCL "world" is completely different from the host language. It looks all good, but it doesn't seem like anyone has really tested ROCm out. There are some benchmarks over at Phoronix testing out their OpenCL implementation on ROCm (and apparently, the ROCm OpenCL is a different compiler from the older AMD stuff). -------- Its kind of hard for me to see where AMD's "focus" is. If I had to bet, it seems like ROCm HCC has the most effort right now. In the past two years, ROCm HCC has seen multiple releases from 1.0 through 1.7, and their Github page indicates a lot of activity. Way more activity than AMD's OpenCL forums for example. But there's no community forums of ROCm / HCC yet! I think most discussions happen in GitHub issues, which is... a subpar community solution. AMD has been working on a lot of technologies that seem redundant: Vulkan Compute Shaders, OpenCL Compute, and now ROCm / HCC Compute. In addition, there's Microsoft C++ AMP and Microsoft DirectCompute. NVidia has done a good job at communicating "CUDA" as the #1 platform of choice on NVidia systems. But I'm not entirely sure which technology AMD is focusing on right now. If I had to guess, they're all in on ROCm / HCC. But clarifications would be nice. I do realize that AMD is also offering OpenCL support on ROCm. But things like Packed FP16 are better supported on ROCm HCC, and I don't think packed FP16 is supported on ROCm OpenCL yet. This suggests that AMD's #1 effort is HCC, despite the relative silence on the HCC platform. See this ROCm issue for example: https://github.com/RadeonOpenCompute/ROCm/issues/219 https://github.com/RadeonOpenCompute/ROCm/issues/219

dis-sys 9y ago

there is always such competition - E5-2600v4 vs LGA3647 based processors betting on AMD is nothing different from waiting for an open sourced version of Windows from Microsoft.

dragontamer 9y ago

ROCm looks good, but so did OpenCL and I'm not very convinced that OpenCL is a good platform to build code in. ROCm addresses a lot of my concerns on paper, but there isn't much buzz on ROCm... specifically HCC. It looks like ROCm builds off of Microsoft's C++ AMP specification (which also looks like a great idea, but that project unfortunately looks dead). So it provides a "unified" C++ source code that can compile down into x86 OR GPU Assembly (GCN in the case of AMD). Developing OpenCL requires a bit of "copy/paste" going on whenever you start processing data on the GPU vs CPU. Because the OpenCL "world" is completely different from the host language. It looks all good, but it doesn't seem like anyone has really tested ROCm out. There are some benchmarks over at Phoronix testing out their OpenCL implementation on ROCm (and apparently, the ROCm OpenCL is a different compiler from the older AMD stuff). -------- Its kind of hard for me to see where AMD's "focus" is. If I had to bet, it seems like ROCm HCC has the most effort right now. In the past two years, ROCm HCC has seen multiple releases from 1.0 through 1.7, and their Github page indicates a lot of activity. Way more activity than AMD's OpenCL forums for example. But there's no community forums of ROCm / HCC yet! I think most discussions happen in GitHub issues, which is... a subpar community solution. AMD has been working on a lot of technologies that seem redundant: Vulkan Compute Shaders, OpenCL Compute, and now ROCm / HCC Compute. In addition, there's Microsoft C++ AMP and Microsoft DirectCompute. NVidia has done a good job at communicating "CUDA" as the #1 platform of choice on NVidia systems. But I'm not entirely sure which technology AMD is focusing on right now. If I had to guess, they're all in on ROCm / HCC. But clarifications would be nice. I do realize that AMD is also offering OpenCL support on ROCm. But things like Packed FP16 are better supported on ROCm HCC, and I don't think packed FP16 is supported on ROCm OpenCL yet. This suggests that AMD's #1 effort is HCC, despite the relative silence on the HCC platform. See this ROCm issue for example: https://github.com/RadeonOpenCompute/ROCm/issues/219 https://github.com/RadeonOpenCompute/ROCm/issues/219