AMD Releases Open-Source LLVM Backend for AI Engine Processors

TapTechNews June 11th, AMD releases open-source LLVM backend Peano specifically developed for AMD/Xilinx AI Engine processors, supporting chips including several Ryzen APUs.

 AMD Releases Open-Source LLVM Backend for AI Engine Processors_0

According to the introduction, the current project mainly supports the AIE2 architecture implemented by the XDNA accelerator in the Phoenix and HawkPoint processors; the support for the StrixPoint processor with the XDNA2 architecture NPU will come later.

Ryzen 7040 series Phoenix processor: XDNA architecture NPU, computing power up to 10 TOPS (10 trillion operations per second).

Ryzen 8040 series HawkPoint processor: XDNA architecture NPU, computing power up to 16 TOPS.

Ryzen AI 300 series StrixPoint processor: XDNA2 architecture NPU, computing power up to 50 TOPS.

To support the unusual architectural features of the AIE Engine, the project has added LLVM support for some specific features:

 AMD Releases Open-Source LLVM Backend for AI Engine Processors_1

In addition, support for Clang, LLD, binutils, Compiler-RT, and LLVM-LIBC is also included.

We hope to invite the community to comment on these methods and begin to promote these general improvement measures upstream.

TapTechNews noted that the Peano team is currently actively improving OCR and supporting the latest version of the AIE structure, especially the XDNA2 accelerator in the StrixPoint processor, which will be released in July.

Likes