Triton

Type: Product (open-source kernel language and compiler; the default target of PyTorch's own code generator, TorchInductor)

Triton lets you write a GPU kernel in Python and leaves the chip-specific work to the compiler. On this rung it is the portable baseline that hand-written kernels are measured against — and lose to: FlashAttention-4 reports up to 2.7× over Triton on B200 (FlashAttention-4).

The January 2026 design note from Meta's Triton team describes the compiler closing part of that gap by itself. Automatic warp specialization (autoWS) splits a kernel's loop into warp groups with separate jobs — loading, matrix maths, writing results — and plans the buffers and synchronisation between them. On B200 flash-attention forward, the authors report 1.5–2× over stock Triton, close to Gluon and cuDNN, with cuDNN still 10–20% ahead (Triton autoWS). It is experimental, works best on attention forward today, and supports Hopper and Blackwell only; AMD support is on the roadmap.

Why the source is graded as it is

The autoWS post is written by the team shipping the feature, with one performance sentence and no configurations — vendor-authored, graded weak to moderate for the speedup and moderate for the design description. The two Triton numbers on this rung (2.7× behind FlashAttention-4; 1.5–2× gained by autoWS) come from different kernels and baselines and must not be chained into one figure.

Mentioned In

• Kernel portability — the portable-versus-hand-tuned gap

Related Entities

• FlashAttention — the hand-written kernel Triton is measured against
• NVIDIA — cuDNN, the vendor library autoWS still trails