LLM / CPU research
Mout CPU Research
GitHubA good hypothesis must pass all three gates.
A4
G128
G128
93.5212%Q4 gap closed · 4.1250 ATTENTION_Q BPW
Runtime: 42.83% / 43.85% of Q4 (n128 / n1024). Production winner: none.
Measure. Validate. Decide.
Low-bit CPU inference research where quality, model size, and AVX2 runtime were evaluated as one decision—not isolated attractive metrics.
A4-G128-SYM achieved the best tested quality/size research result. Its AVX2 runtime did not meet the joint production target. Research v1 is complete: Q4 remains the production reference, with no custom format winner.
01
Representation
Low-bit formats and quantization
02
Parity
Scalar / packed / AVX2 verification
03
Decision gate
Quality + size + runtime
What is verified.
Research outcome
- Q4 remained the production reference; custom format production winner: none.
- A4-G128-SYM: 93.5212% Q4 gap closed, 4.1250 ATTENTION_Q BPW.
- Runtime mismatch: 42.83% / 43.85% Q4-relative throughput—research result, not a production win.
Basis for the claims
Frozen reports, source snapshots, parity utilities, manifests, and raw logs in the archive.
The exact final decision table preserves quality/size/runtime context.
Research v1 / final report