Vulkan 1.4.352 is out today as the latest minor spec update to this Khronos API. Besides just a few fixes/clarifications, there is one new extension and that is a NVIDIA vendor extension for cooperative matrix decode vector support.

VK_NV_cooperative_matrix_decode_vector is the new extension developed by NVIDIA and is for extending the decode callback added previously in VK_NV_cooperative_matrix2 to now support decoding multiple matrix elements per invocation. With most quantized weight formats being unpacked in groups, it's more efficient and allows for better optimizations being able to unpack multiple elements in one go rather than having to call the decode function once per element. Overall a nice improvement for NVIDIA's cooperative matrix support in Vulkan that helps with machine learning ambitions in the Vulkan API.

[1]

The VK_NV_cooperative_matrix_decode_vector introduction the main highlight of today's [2]Vulkan 1.4.352 spec update.

NVIDIA already published a new [3]Vulkan beta driver for Windows and Linux that rolls out this new cooperative matrix decode vector extension. The beta driver is 596.54 for Windows and 595.44.08 for Linux.



[1] https://www.phoronix.com/image-viewer.php?id=2026&image=nvidia_vector_lrg

[2] https://github.com/KhronosGroup/Vulkan-Docs/commit/e2843a23d3c5a2c5e93e27d3b0fcd2c1087b0b64

[3] https://developer.nvidia.com/vulkan-driver