SCALE 1.3 Adds BFloat16 & Other New Features For Compiling CUDA Apps On AMD GPUs
A new software project covered on Phoronix last year was SCALE for natively compiling CUDA applications for AMD GPUs. This "clean room" implementation of CUDA building off the open-source LLVM codebase continues going strong and out this week is SCALE 1.3 with more features and hardware support for compiling CUDA software for AMD GPU execution.
SCALE continues to be developed by Spectral Compute and completely independent from AMD. SCALE's goal remains being a GPGPU programming toolkit for natively compiling CUDA software for AMD GPUs while not requiring modifications to the CUDA program itself. This nvcc-compatible compiler and other elements of SCALE and its runtime continue improving for enhancing the possibilities of CUDA software built for AMD GPUs.
With the new SCALE 1.3 release they shifted their codebase from being built atop LLVM 17 to now on LLVM 19. They also added support for ROCm 6.3.1 upstream.
New hardware support includes the Radeon RX Vega "GFX902" graphics processors. Additionally, an enterprise edition of SCALE now supports the AMD Instinct accelerators with the GFX908, GFX90A, GFX940, and GFX942 targets.
Some of the other changes to SCALE 1.3 include a new "scaleenv" environment command to make it easier to use SCALE, support for BFloat16 instructions, support for simulating a warp size of 32, compatibility improvements for non-CMake build systems, and various library improvements. The SCALE library has new multi-GPU APIs supported and various other additions. The NVCC-compatible compiler has also seen various compatibility enhancements.
SCALE 1.3 also now provides official Rocky Linux 9 packages alongside their existing Ubuntu packages and their generic tarball.
Those wishing to learn more about the SCALE 1.3 software release can do so via SCALE-Lang.com.
SCALE continues to be developed by Spectral Compute and completely independent from AMD. SCALE's goal remains being a GPGPU programming toolkit for natively compiling CUDA software for AMD GPUs while not requiring modifications to the CUDA program itself. This nvcc-compatible compiler and other elements of SCALE and its runtime continue improving for enhancing the possibilities of CUDA software built for AMD GPUs.
With the new SCALE 1.3 release they shifted their codebase from being built atop LLVM 17 to now on LLVM 19. They also added support for ROCm 6.3.1 upstream.
New hardware support includes the Radeon RX Vega "GFX902" graphics processors. Additionally, an enterprise edition of SCALE now supports the AMD Instinct accelerators with the GFX908, GFX90A, GFX940, and GFX942 targets.
Some of the other changes to SCALE 1.3 include a new "scaleenv" environment command to make it easier to use SCALE, support for BFloat16 instructions, support for simulating a warp size of 32, compatibility improvements for non-CMake build systems, and various library improvements. The SCALE library has new multi-GPU APIs supported and various other additions. The NVCC-compatible compiler has also seen various compatibility enhancements.
SCALE 1.3 also now provides official Rocky Linux 9 packages alongside their existing Ubuntu packages and their generic tarball.
Those wishing to learn more about the SCALE 1.3 software release can do so via SCALE-Lang.com.