Huawei open-sources AscendNPU IR for Triton and Ascend 950

A compiler architect took the stage at HyperAI's technical salon on August 1 with a problem that usually stays hidden from users: how to turn a few lines of tensor code into work distributed across memory, cores and on-chip pipelines. Hai Lijuan, Huawei's AscendNPU IR architect, presented the open-sourced compiler layer that connects Triton and other languages to Ascend hardware, including the Ascend 950.

Triton lowers the barrier for writing AI operators with Python-like syntax and block-level tile programming. Developers describe how data is divided and processed; the compiler can automate memory coalescing and shared-memory management. AscendNPU IR takes that model toward the chip. Built on MLIR, it links high-level languages above to LLVM IR below, then compiles the result into Ascend binaries.

The design separates what can work across hardware from what must follow a particular chip. HFusion provides a hardware-independent layer for multi-dimensional fusion optimization. HIVM then maps logical tensors onto physical memory, assigns cube and vector operations to their respective cores, and enables vectorization for high-performance instructions. The stack covers Ascend architectures from A2/A3 through Ascend 950.

The Ascend 950 brings a deeper change inside that mapping layer. HIVM expands from memory-based SIMD to register-based SIMD and SIMT, while cube and vector cores can interact through tightly coupled on-chip communication instead of exchanging data through global memory. New compiler features address complex control flow, nested loops, flash-attention-style pipelining and a mode in which two vector cores each compute half the data against one cube core.

So what, concretely? Developers can reuse more familiar programming approaches while targeting Ascend chips, rather than rebuilding every operator around low-level memory and instruction details. The code is now open for co-construction through Triton-Ascend and AscendNPU IR. For teams without an Ascend environment, Huawei's community offers HiDevLab, with 100 free hours of compute time. The open-source release does not by itself prove performance in independent workloads, but it gives outside developers a place to test, adapt and improve the compiler path.