Vulkan 1.4.352 arrives with NVIDIA matrix decoding

Khronos has released Vulkan 1.4.352, a minor update that brings the major new feature of the VK_NV_cooperative_matrix_decode_vector extension. This extension, signed by NVIDIA, allows decoding multiple matrix elements per invocation, expanding the capabilities of VK_NV_cooperative_matrix2. Corrections and clarifications for the API are also included.

Vulkan API pipeline visualization, cooperative matrix decode process, NVIDIA GPU die with parallel vector units actively decoding multiple matrix elements per invocation, blue and orange glowing data streams flowing through cooperative matrix hardware blocks, shader core clusters processing decode vectors simultaneously, technical illustration style, clean engineering render, metallic silicon surface with microscopic circuit detail, cool blue ambient light with warm orange data paths, precise architectural diagram aesthetic, ultra-detailed semiconductor visualization

Vector Decoding for Cooperative Matrices 🚀

The VK_NV_cooperative_matrix_decode_vector extension optimizes working with matrices in shaders. While the previous extension allowed decoding a single element per invocation, this new version processes entire vectors, reducing the number of required calls. This benefits workloads such as neural network inference or data compression, where access to data blocks is frequent. The implementation is specific to NVIDIA hardware and requires support from recent drivers.

NVIDIA and Its Obsession with Decoding Every Last Byte 🔥

NVIDIA doesn't rest: if before you decoded one element at a time, now you can do it in bursts. Soon we'll see an extension to decode the entire matrix in one go, and then another so the GPU can make you coffee while you wait. Meanwhile, AMD and Intel developers look on askance, wondering if this is magic or just another way to sell more graphics cards.