Introduced at the end of March was Vulkan 1.4.311 with VK_KHR_shader_bfloat16 for allowing BFloat16 “BF16” operations within SPIR-V shaders with SPV_KHR_bfloat16. This BFloat16 support can be beneficial for Vulkan machine learning / AI workloads and other use cases moving forward. Now the Mesa Radeon Vulkan driver “RADV” is the newest wiring up support for VK_KHR_shader_bfloat16.
Since the Vulkan 1.4.311 release there was a quick NVIDIA Vulkan driver beta with support and at the end of April the Intel ANV driver added BFloat16 support for Mesa 25.2. Now in following that Intel ANV work and changes to the common SPIR-V code for BF16, the RADV implementation was merged today to Mesa Git.
Georg Lehmann who worked on this VK_KHR_shader_bfloat16 for the open-source Radeon Vulkan driver noted though that AMD RDNA3 (GFX11) graphics processors have precision issues with BF16 and thus not enabled for now. So for the moment at least the BF16 support with Radeon GPUs is only being exposed for RDNA4 (GFX12) GPUs.
This merge is what landed for this RADV VK_KHR_shader_bfloat16 support. So far AMDVLK as AMD’s official open-source Vulkan API driver for Linux systems isn’t yet exposing this BFloat16 extension.
