Radeon ROCm 3.5 Released With New Features But Still No Navi Support
ROCm 3.5 was released today as the successor to ROCm 3.3 with no v3.4 milestone having been made public. Highlights of ROCm 3.5 include:
- The Heterogeneous Compute Compiler (HCC) has been deprecated in favor of the HIP-Clang compiler for compiling HIP programs. The HIP-Clang code has been seeing a lot of work upstreamed into LLVM/Clang os overall this should be good in the long-run.
- The AMD ROCm OpenCL runtime now supports OpenCL 2.2.
- The AMD ROCm GNU Debugger (ROCgdb) as their derived GDB solution for debugging ROCm GPU and CPU code. There is also a new ROCm Debugger API Library.
- Updates so the AMD RCCL is compatible with the NVIDIA Communications Collective Library NCCL 2.6.4 release.
- A new SMI event interface and library.
- A new API for allowing applications to select the proper memory node for a given accelerator.
- The new Radeon Performance Primitives Library for HIP and OpenCL usage.
- Many bug fixes.
The officially supported GPUs/accelerators are the GFX8 Fiji / Polaris 10 hardware as well as GFX9/Vega hardware. There is also unofficial support for the GFX8 Polaris 11/12 chips as well as GFX7 Hawaii GPUs. While next month marks one year since the launch of the first consumer Navi graphics cards, surprisingly AMD still doesn't have ROCm wired up for GFX10/Navi as of ROCm 3.5.
Download ROCm 3.5 via GitHub.
100 Comments
