Matrix multiplication kernels for Blackwell with Pallas in JAX
The JAX documentation explains how to write custom matrix multiplication kernels for NVIDIA Blackwell GPUs using Pallas, focusing on performance optimization.
By Wendelmaques Pereira · Updated today
Source: Escrevendo kernels de multiplicação de matrizes de alta performance para Blackwell com… (docs.jax.dev). Text prepared with AI from this source.
What happened and what to do
The JAX documentation page explains how to write high-performance matrix multiplication (matmul) kernels for NVIDIA Blackwell GPUs using Pallas, JAX's language for custom kernels. The post that shares the link summarizes the topic as useful for kernel optimization.
For a company running inference or training on Blackwell GPUs, custom kernels can reduce bottlenecks in operations that standard implementations do not exploit well on the hardware. In practice, this requires benchmarking the model to identify the dominant operations, implementing the kernels in Pallas, validating numerical results against a reference, and integrating them into the serving pipeline, with monitoring of latency and cost per token.
How the consultancy can help
Diagnosis of where the model loses performance on Blackwell GPUs, with profiling of the dominant operations. Then implementation and validation of custom kernels in Pallas, with numerical equivalence tests and operation within the client's own environment.
Next step
If your team runs models on Blackwell GPUs and wants to assess gains from custom kernels, send a short description of your case and receive a scoped proposal from Wendelmaques.
Let's assess your project
I reply with scope and a proposal.