Skip to main content

Lab Notebook · Awesomekernel — The Megakernel

Awesomekernel — one persistent GPU kernel, treated as a phase-locked oscillator

A clean-room substrate where a single fused, persistent CUDA kernel owns an entire training step — matmul, attention, MLP, loss, gradient — without ever returning to the host between phases. The target is 25ms/step against a 278ms PyTorch baseline; that target is not yet fully measured end to end, and we say so. What's public here is the thesis and the honesty mechanism that keeps it accountable, not the kernel.

JR
Jon RossFounder, vocabotics — 15 years building safety-critical systemsLab report · dated 30 April 2026
Verified by a human. Drafted with AI, verified by a human. Jon Ross, 30 Apr 2026
Living document. Reviewed 30 Apr 2026
Entry date
30 April 2026
Category
Inference
Lead over the world
the discipline, not the lead
Access
🔒 Protected
ResearchNot run

One kernel, never leaving the GPU for an entire training step — reframed as a phase-locked oscillator, with a counter that catches it if a fused stage silently does nothing.

25ms (target)
target ms/step vs a 278ms PyTorch baseline
936 GB/s
HBM bandwidth on the target RTX 3090
REACH counter
the wire-in discipline that proves each fused stage actually ran
280 files / 9.6MB
scale of the substrate, active late Apr-May 2026

Honest evaluation

Not run

The REACH-counter discipline is real and running; the 25ms/step target itself has not yet been measured end to end.

What we learned — a preview

The honest edges are open even when the engine is not. The full report — what it was, what we built, and the measured internals — is protected.

  • The 25ms/step figure is a target, not yet a fully measured end-to-end result — say so plainly rather than presenting it as achieved.
  • This is core substrate IP; implementation and kernel internals are deliberately kept out of this report.
  • A single fused kernel that never returns to the host is a genuinely unusual design; unusual designs carry real risk of subtle correctness bugs, which is exactly what the wire-in discipline exists to catch.
Protected report

The deep body of “Awesomekernel — one persistent GPU kernel, treated as a phase-locked oscillator” is behind access.

We open the demos, the specs, and the method; we protect the engines, the model internals, and anything that touches commercial, safety-critical work. This report describes an engine — so its details are gated, and no client, company, or project is named. Members can read it in full; if you have a genuine reason to see it, tell us who you are and why.

Proofs & sparks

We demonstrate rather than assert. Each ✅ proof is a visible result with a hard figure.

  • Scope-restricted cooperative-grid kernelclean vs CUTLASS / FA-3a scope-restricted persistent cooperative-grid kernel with a structural fusion criterion cleanly differentiated from CUTLASS / FlashAttention-3 — wired in and council-ratified.
  • Persistent mega-kernel OSconceptual ancestor · partiala persistent mega-kernel OS with ring-buffer syscall dispatch — partially realised via later wave-kernel work; most sibling claims remain dormant, stated plainly.
  • Fused training megakernels — headline mechanismsH4 gate-passeda compile-time A→B scaffold, provenance-tagged training tape, embed-prefix-fused forward and block-scheduled backward megakernels — H4 passes the gate; H3 is design-ratified only.
  • Curriculum-order + fusion spike-absorption0 spikes vs 4–5 baselinecurriculum-order and fusion spike-absorption — empirically validated and gate-passed: 0 loss-spikes versus 4–5 in the baseline.

Where this connects

    We use cookies.