PORTABILITY IS A MYTH: WHY THE BEST AI STACKS WILL NEVER BE HARDWARE-AGNOSTIC

Summary from the TLDR AI newsletter

AI kernel portability is structurally impossible because TPU's Pallas, NVIDIA's CuTile and CUTLASS, AWS's NKI, AMD's FlyDSL, and Tenstorrent's tt-Metalium each expose hardware-specific concepts that no universal DSL can unify. The evidence: MaxText's MoE grouped matmul ships as 282 lines of Pallas on TPU while flashinfer's equivalent for Blackwell SM100 takes 4 million lines of generated CUDA, with zero shared code because the algorithms themselves diverge across hardware.