Technical Architecture and Systematic Benchmarking of the OpenXLA Ecosystem
DRAFT v2 — April 2026
A cross-platform analysis of modern ML compilers: an architectural study of the OpenXLA pipeline (CHLO → StableHLO → XLA → LLVM) paired with a proposed 3×3 benchmarking protocol evaluating three workloads (LLM inference, dense training, sparse embedding training) across TPU v6e, NVIDIA H200, and AMD MI300X.
Status: Proposed research protocol. Architectural analysis and methodology are complete; experimental results are forthcoming pending hardware access.