Skip to content

Releases: IR14/gf137-edge-kernel-replication

Release list

v0.1.0 - GF(137) Edge-Kernel Replication Baselines

v0.1.0 - GF(137) Edge-Kernel Replication Baselines Latest
Latest

Choose a tag to compare

View all tags
@IR14 IR14 released this 31 May 00:49
v0.1.0
7c61fc1

Release Notes

v0.1.0 - GF(137) Edge-Kernel Replication Baselines

Release date: 2026-05-31

This is the first release-ready replication package for the GF(137)
edge-kernel audit.

Included Tracks

Track Description Primary report
HYP-002 Single-shape GF(137) replication against equivalent float32 modular baselines outputs/edge_kernel_replication.md
HYP-003 Three-shape sweep against float32 modular and plain uint8_t controls outputs/quantized_baseline_sweep.md
HYP-004 Hand-written C++ int8 dense baseline with int32 accumulation outputs/industrial_int8_baseline.md
HYP-005 ONNX Runtime CPU MatMulInteger int8 external-runtime baseline outputs/onnxruntime_int8_baseline.md

External Replication Artifacts

Environment Reports
Apple ARM local outputs/edge_kernel_replication.md, outputs/quantized_baseline_sweep.md, outputs/industrial_int8_baseline.md, outputs/onnxruntime_int8_baseline.md
Linux VPS x86_64 outputs/vps_vds2640757_expanded_edge_kernel_replication.md, outputs/vps_vds2640757_hyp003_quantized_baseline_sweep.md, outputs/vps_vds2640757_hyp004_industrial_int8_baseline.md, outputs/vps_vds2640757_hyp005_onnxruntime_int8_baseline.md
GitHub Actions Ubuntu HYP-002, HYP-003, HYP-004, and HYP-005 workflow runs linked from README.md and docs/gf137_edge_kernel_note.md

Supported Claim

The artifacts support a narrow engineering claim:

  1. GF(137) model storage is 4x smaller than equivalent float32 modular
    storage for the tested dense kernel layouts.
  2. Equivalent GF(137) rows match the frozen reference outputs exactly in the
    tested runs.
  3. GF(137) is consistently faster than equivalent C++ float32 modular
    arithmetic in the tested HYP-002 and HYP-003 layouts.
  4. Against standard int8-style inference, speed is shape- and
    hardware-dependent. ONNX Runtime int8 is faster than GF(137) on 2/3
    tested shapes in both Apple ARM and Linux VPS runs.

Explicit Non-Claims

This release does not claim that:

  • GF(137) is faster than all int8 inference implementations;
  • GF(137) replaces ONNX Runtime, TFLite, or vendor-optimized kernels;
  • the result scales to large neural networks;
  • the benchmark establishes any fundamental-physics result.

Reproduction

Run the full local audit:

./scripts/run_hyp002.sh
python scripts/assert_hyp002_pass.py

./scripts/run_hyp003.sh
python scripts/assert_hyp003_pass.py

./scripts/run_hyp004.sh
python scripts/assert_hyp004_pass.py

./scripts/run_hyp005.sh
python scripts/assert_hyp005_pass.py

For older Linux x86_64 CPUs whose NumPy wheels may lack the required CPU
baseline, use:

NUMPY_SPEC="numpy==1.26.4" ./scripts/run_hyp002.sh
NUMPY_SPEC="numpy==1.26.4" ./scripts/run_hyp003.sh
NUMPY_SPEC="numpy==1.26.4" ./scripts/run_hyp004.sh
NUMPY_SPEC="numpy==1.26.4" ONNX_SPEC="onnx" ONNXRUNTIME_SPEC="onnxruntime" ./scripts/run_hyp005.sh

Next Work

The next audit should add a TFLite or TFLite Micro baseline. That should be
treated as a boundary test, not as an attempt to preserve the GF(137) speed
claim.

Assets 2
Loading