A small research project that grew out of debunking one hype claim — "embed a 256 KB neural model in L2/L3 cache and binary translation becomes near-free" — and salvaging the parts that are actually real.
The original "Neural Binary Translation Engine" pitch fails on two hard facts:
- You can't pin a buffer into L2/L3 from userspace. Hot data stays statistically resident via LRU; it is not "locked," and the workload you're translating evicts it.
- Inference is compute-bound, not reach-bound. Keeping weights in cache fixes memory latency. A forward pass is a flood of MACs — the bottleneck is ALU throughput, which cache residency does nothing for. (mechanism 1 measures this on real hardware.)
But three salvageable mechanisms fall out of taking the idea apart. This repo is the central home for all three. Two of them are worth shipping as standalone tools; the third stays here as an integration/teaching harness.
| Component | Role | Destiny |
|---|---|---|
residency-bench (mechanism 1) |
Cache/working-set residency + bottleneck classifier microbench | → standalone repo A (generally useful, publish) |
jit-oracle (mechanism 3) |
Offline-trained Box64 eager-translate model, distilled to a static profile — zero inference in the hot loop | → standalone repo B (the novel bit, publish) |
minidbt (mechanism 2) |
Toy translate-once dispatch loop; the educational core of a DBT | stays central (don't reinvent Box64) |
The translator target is Box64 (MIT, the Winlator default). The two standalone repos can be vendored back here as submodules so this repo stays the single source of truth, the writeup home, and the place the fork-and-gut test harness lives.
Source-available, no-redistribution — see LICENSE. These repos are
public so the work can be read and evaluated, but they are not open-source:
no copying, running, redistribution, or derivative works without written
permission. (A public Git host technically permits git clone; the no-download /
no-use restriction is enforced by the license terms, not a platform setting.)
Third-party components referenced or patched (Box64, Wine, DXVK, …) keep their own
licenses.
docs/SPEC.md— full spec for all three mechanismsdocs/FEASIBILITY.md— scored feasibility + what's missingdocs/PHASES.md— measure-first phased roadmapdocs/ETHICS.md— what we will and won't builddocs/TESTING.md— the fork-an-open-Winlator, gut-and-replace methodology
Spec/feasibility phase, Phase 0. minidbt (mechanism 2) lives here under
components/. Mechanism 1 graduated to its own repo (residency-bench);
mechanism 3 (jit-oracle) is a gated, spec-stage scaffold. Next concrete step:
get residency-bench building + an aarch64-Android target, run it on the S24U.