VGX-01 Flagship · Founder VERIFIED ← Return to console

Silicon Logic

Benchmarks you can verify, not just believe.

Independent local AI inference testing with signed artifacts, public methodology, and reproducible numbers.

01 What it is

Silicon Logic measures how AI models actually run on local hardware — tokens per second, latency, power — and publishes the results as editorial work a reader can check.

02 Why it exists

Local AI inference is full of numbers and short on defensible ones. Hardware vendors publish competing claims for the same silicon. Community benchmarks are useful but unauditable. Video reviews run impressive tests with methodology no one can re-run. If you’re choosing hardware for local inference, you can find a thousand charts — and almost none you could defend in front of someone who disagrees.

03 How it works

The measurement plan is committed before the runs happen. Every run produces a signed artifact — canonical bytes under an Ed25519 signature — and the raw outputs ship alongside the headline numbers. The methodology is versioned and published as part of the work, not a footnote under it. The statistics are built for noisy machines: repeated trials, with variance reported, not hidden.

04 What it measures

Throughput, time to first token, latency percentiles, memory pressure, and measured power — wall and SoC — with thermals observed across sustained runs. On hardware readers actually own: Apple Silicon (a MacBook Pro M5 Max with 36 GB) and consumer GPUs (RTX 5080), across the runtimes people actually use — llama.cpp, MLX, Ollama.

05 The role

I founded Silicon Logic and build all of it: the measurement pipeline, the schema that keeps the Python, TypeScript, and database layers honest with each other, and the publication site that renders the artifacts.

06 Proof — benchmark certificate VERIFIED · 2026-06-10

id 01KS19XRVEMDPQKCP77C9GCCYG model qwen 2.5 7b instruct · q4_k_m gguf runtime llama.cpp b9070 scenario interactive throughput 83.2159 tok/s ttft 20 ms power 12.1046 w avg · 6.8747 tok/w ran 2026-05-19

signature ed25519 · valid sig/XXPiTNVgGnWIT7Df4pm1mjdc0+ssaO6AvxfVqxAVFkDI4l6nDJpcpXvXDgB7ZX2GtNClaOHfWJ37VA22jLwBw== signeryRZYrqoqwd0VPwKxTyF24z1IMW1AdIzgXyVSfyCt5WQ= sha25648d35cd9413be6fb2475b3b95261620d45fb418c99016ad0ef9463348895f73e

07 Check it yourself

Don’t trust this page — check it.

The procedure below is the exact verification this record passed, run against the repository on 2026-06-10.

# Silicon-Logic/silicon-logic · services/pipeline$ uv run python>>> import json>>> from pipeline.schemas.benchmark_run import BenchmarkRun>>> from silicon_logic.signing import verify_benchmark_run>>> raw = json.load(open("data/runs/01KS19XRVEMDPQKCP77C9GCCYG.json"))>>> verify_benchmark_run(BenchmarkRun.model_validate(raw))True

The signer key’s SHA-256 fingerprint — 48d35cd9…f73e — matches the maintainer fingerprint published in the Silicon Logic repository README.

Read Silicon Logic → X @siliconlogic_ GitHub Silicon-Logic