Decoupling Compute and Memory for Async GPUs

https://news.ycombinator.com/item?id=48676372
by yiyingzhang • 28 days ago
8 2 28 days ago

Cool open-source project that introduces a new programming model for decoupling compute and memory for NVIDIA GPUs that supports asynchronous memory operations (e.g., Hopper). 12% perf improvement over SOTA and 67% less kernel code.

Paper: "VDCores: Resource Decoupled Programming and Execution for Asynchronous GPU" arXiv:2605.03190

Related Stories

Loading related stories...

Source preview

news.ycombinator.com