|
Hi, I noticed that the kernels in this tutorial mainly target Blackwell GPUs such as B200. If I don't currently have access to a Blackwell GPU, how much of the tutorial can I still follow meaningfully? For example, are the concepts around layouts, pipelining, and GEMM optimization still useful to practice on older NVIDIA GPUs, even if some features like TMEM or newer TMA instructions are unavailable? Thanks! |
Replies: 1 comment
|
Hi @H1sk897 Thanks for asking the question! Yes. Many of the concepts here, such as tiling, layouts, pipelining, and memory hierarchy, are general and not specific to Blackwell. The current examples mainly target Blackwell because they use newer hardware features like TMEM and the latest TMA / tensor core instructions. TIRx itself is not intended to be limited to Blackwell, and we plan to support more backends in the future. |
Hi @H1sk897 Thanks for asking the question!
Yes. Many of the concepts here, such as tiling, layouts, pipelining, and memory hierarchy, are general and not specific to Blackwell.
The current examples mainly target Blackwell because they use newer hardware features like TMEM and the latest TMA / tensor core instructions. TIRx itself is not intended to be limited to Blackwell, and we plan to support more backends in the future.