Evidence, not dates

Build order matters.

Right now

The workload is pinned. Build out execution.

M5 is complete and M6 is active: accelerator nodes 0–217 and the explicit host tail pass tested inputs. Independent review and a command-delivery decision remain before a stable ABI; final corpus release follows review.

In progress

Close the stable ABI

Independent review and a practical command-delivery decision remain before freeze.

In progress

Review conformance

Validate independent fixture derivations and tie the final corpus to the frozen ABI.

In progress

Measure complete command-path accuracy

Build a faster command-equivalent runner, recheck fixtures and six inputs, then evaluate all COCO val2017 images.

Working software →
M0–M12

The complete milestone chain.

M6 proceeds against the explicitly pinned M3/M4 candidate. RTL starts after independent review and stable contract freeze; a complete v0 additionally requires the full workload execution path.

M0
Implemented

Repository foundation

Structure, build entry points, documentation, CI skeleton, and licensing are tracked.

M1
Implemented

Numerical golden model

FP8/INT8 semantics and numerical, layout, activation, and operation tests pass.

M2
Implemented

Functional command simulator

Versioned 128-byte command encoding, memory behavior, errors, and simulator tests exist.

M3
In progress

Freeze the v0 contract

ABI 0.1 is a candidate. Stable freeze awaits independent review and a practical command-delivery decision.

M4
In progress

v0 conformance package

46 independent binary fixtures, schema, registry, integrity checks, runner, and CI are implemented.

M5
Implemented

Pin + calibrate the first YOLO workload

Artifact, graph, preprocessing, host boundary, FLOAT baseline, INT8 calibration, proxy accuracy, and diagnostics are recorded. Exact target execution belongs to M6.

M6
In progress

Minimal compiler + simulated runtime

Accelerator nodes 0–217 and the explicit Python host tail pass synthetic and multi-input comparisons. Next: a faster equivalence-checked runner and COCO-wide command-path accuracy.

M7
Planned

First RTL vertical slice

Decode one command path through INT8 MAC, accumulation, requantization, completion, errors, and differential tests.

M8
Planned

Incremental RTL coverage

Add DMA, convolution, and utility operations one conformance-gated step at a time.

M9
Planned

FPGA selection + v0 bring-up

Select from measured probes and publish reproducible timing, resources, accuracy, transfers, and latency.

M10
Long-term

Second perception workload + reprogramming proof

Run a separately pinned perception graph on the same measured FPGA bitstream.

M11
Long-term

v1 RISC-V + native FP8 research

Add only features justified by measured v0 and reprogramming evidence.

M12
Long-term

ASIC feasibility + test core

Resolve process, SRAM, IO, clock, power, DFT, verification, packaging, and fabrication collateral.

Non-negotiable

Rules that keep results real.

These controls apply across software, compiler, RTL, FPGA, and eventual silicon.

01

One source of truth

Opcodes, fields, layouts, error codes, and ABI versions stay aligned across every consumer.

02

Independent oracles

The golden expectation remains separate from the implementation being tested.

03

No silent fallback

Unsupported operations, host partitions, errors, saturation, and approximations are declared.

04

Reproducible evidence

Pin revisions, tools, commands, seeds, artifacts, models, and measurement conditions.

05

Intermediate comparison

Compare tensors layer by layer instead of trusting only a plausible final detection.

06

Errors are first-class

Negative cases and illegal states receive the same care as successful execution.