What if the physical processor—not merely its architecture—were an input to program specialization?
Portable Program
+
AMD Architecture Knowledge
+
Physical Machine Measurements
↓
Hardware-Native Execution Profile (HNEP)
The runtime loads an HNEP, dispatches cheaply, and falls back to baseline on fingerprint mismatch. Correctness gates and statistical skepticism beat mysterious “AI optimization.”
Not affiliated with, endorsed by, or certified by Advanced Micro Devices, Inc.
Primary evidence path today is a single physical Zen host (e.g. Ryzen 9 9950X). Multi-machine Silicon Split YES/NO is deferred until a second Zen box is available—without blocking professional single-host research.
cargo run -p silicera-cli -- doctor
cargo run -p silicera-cli -- calm-check
cargo run -p silicera-cli -- eval -o out/single-machine-eval.json
cargo run -p silicera-cli -- remarks out/profile.hnep -o out/remarks.yaml
cargo run -p silicera-cli -- about
| Artifact | Role |
|---|---|
| Knowledge pack | Public Zen architecture facts (profiles/amd/) |
| HNEP | Measured winners + size-class decision tree |
| SCF | Compiler feedback JSON |
| Remarks | Experimental LLVM-style YAML |
| Eval pack | One-shot single-Zen report + repro sidecar |
Independent research — dual-licensed MIT OR Apache-2.0. If you use Silicera, you can sponsor via thanks.dev/u/gh/theworker02.