Radeon ROCm 4.3 Released With HMM Allocations, Many Other Improvements

AMD has released ROCm 4.3 as the newest version of their Radeon Open eCosystem stack for providing open-source GPU compute and CUDA portability for their supported graphics processors under Linux. ROCm 4.3 is the biggest update we've seen for this important enterprise piece to their enterprise GPU compute stack in a while.

With ROCm 4.3 there are many changes in store. ROCm 4.3's HIP now supports Heterogeneous Memory Management (HMM) allocations on Linux and will leverage HMM allocations automatically, NUMA-aware host memory allocations, new atomic system scope atomic operations, support for indirect function calls and C++ virtual functions within the ROCm compiler, improved data center tool integration, support for reporting all HBM temperature values per stack, rocBLAS performance optimizations, rocRAND GFX1030 support along with GFX90A and re-enabling GFX803, and many improvements to rocSPARSE, rocSOLVER, and an assortment of other changes across the board.

ROCm 4.3 is officially just supported for CDNA GPUs (AMD Instinct MI100) and GFX9 (Vega 10 / Vega 7nm) graphics processors with unofficial support for Polaris/GFX8 and Hawaii/GFX7 hardware. While there has been signs of progress on GFX10/Navi, there still is not any official support for Radeon RX 6000/7000 series graphics cards with ROCm yet.

More details on the many ROCm 4.3 changes can be found via the project's GitHub.