Radeon ROCm 4.1 Released - Still Without RDNA GPU Support
ROCm 4.0 released back in December with "CDNA" GPU support while now ROCm 4.1 has been released as the newest quarterly feature release to this open-source Radeon compute stack focused primarily on HPC/data-center needs.
ROCm 4.1 delivers on several new features but before anyone asks, no, there still is not any GFX10/RDNA GPU support for either the Radeon RX 5000 or RX 6000 series. The ROCm compute support remains focused on Vega GPUs and CDNA GPUs, the AMD Instinct MI100. There remains full-support-but-not-guaranteed coverage for Polaris and Hawaii GPUs. Those with the newer RDNA consumer GPUs wanting to use the Radeon Open eCosystem stack, you are left waiting still with ROCm 4.1. The ongoing delay in RDNA GPU support for ROCm appears to be due to AMD's focus on getting everything up to par with the CDNA GPU support with forthcoming super-computer deployments and other big ticket HPC customers.
What there is new with ROCm 4.1 is TargetID handling to allow for multiple configurations per target, new tools of "roc-obj-ls" and "roc-obj-extract" for listing and extracting ROCm code objects within HIP executables and shared objects, Grafana integration in the ROCm Data Center Tool, expanded kernel coverage in rocSPARSE, new cooperative group functions for HIP to match CUDA, and a variety of OpenMP enhancements in offloading to AMD GPUs.
ROCm 4.1 was released just a few minutes ago over on GitHub.
ROCm 4.1 delivers on several new features but before anyone asks, no, there still is not any GFX10/RDNA GPU support for either the Radeon RX 5000 or RX 6000 series. The ROCm compute support remains focused on Vega GPUs and CDNA GPUs, the AMD Instinct MI100. There remains full-support-but-not-guaranteed coverage for Polaris and Hawaii GPUs. Those with the newer RDNA consumer GPUs wanting to use the Radeon Open eCosystem stack, you are left waiting still with ROCm 4.1. The ongoing delay in RDNA GPU support for ROCm appears to be due to AMD's focus on getting everything up to par with the CDNA GPU support with forthcoming super-computer deployments and other big ticket HPC customers.
What there is new with ROCm 4.1 is TargetID handling to allow for multiple configurations per target, new tools of "roc-obj-ls" and "roc-obj-extract" for listing and extracting ROCm code objects within HIP executables and shared objects, Grafana integration in the ROCm Data Center Tool, expanded kernel coverage in rocSPARSE, new cooperative group functions for HIP to match CUDA, and a variety of OpenMP enhancements in offloading to AMD GPUs.
ROCm 4.1 was released just a few minutes ago over on GitHub.
55 Comments