Vulkan 1.4.356 is out today and it's interesting for the lone new extension debuting: VK_EXT_shader_ocp_microscaling_types. The VK_EXT_shader_ocp_microscaling_types is for enabling the Open Compute Project's Microscaling MX data types to help with machine learning workloads with Vulkan.
The Open Compute Project Microscaling Formats (MX) include MXFP4, MXFP6, MXFP8, and MXINT8 for improving the efficiency of AI training and inference. VK_EXT_shader_ocp_microscaling_types adds support for the floating-point types defined in the OCP MX format specification but does not provide any direct support for microscaled tensors or operations on microscaled data.
Industry demand has led to this extension being worked on to support these newer data types for increased compute throughput and lower memory bandwidth requirements. NVIDIA, Samsung, Imagination, and Arm contributed to the VK_EXT_shader_ocp_microscaling_types extension that has been in the works the past several months.
The introduction of VK_EXT_shader_ocp_microscaling_types is the main highlight of today's
The Open Compute Project Microscaling Formats (MX) include MXFP4, MXFP6, MXFP8, and MXINT8 for improving the efficiency of AI training and inference. VK_EXT_shader_ocp_microscaling_types adds support for the floating-point types defined in the OCP MX format specification but does not provide any direct support for microscaled tensors or operations on microscaled data.
Industry demand has led to this extension being worked on to support these newer data types for increased compute throughput and lower memory bandwidth requirements. NVIDIA, Samsung, Imagination, and Arm contributed to the VK_EXT_shader_ocp_microscaling_types extension that has been in the works the past several months.
The introduction of VK_EXT_shader_ocp_microscaling_types is the main highlight of today's