T27.AI

Measured evidence

Every number here was measured.

Hardware claims are cheap to make and hard to check, so this page collects the results behind everything else: what was built, what it measured, and how it was verified. Where something is a submission rather than a win, or a prototype rather than a product, it says so.

Results

GF16 4×4 matmul — 32,252 LUT with zero hard multipliers

A 4×4 matrix multiplier over my own GF16 format, synthesised for Artix-7. It maps into fabric with no DSP48 blocks at all, or 21,223 LUTs if the 64 hard multipliers are allowed. The block is combinational — no registers, so no clock and no frequency figure belongs to it.

100% held-out — a network that trains on the FPGA

Forward pass, gradient and weight update all in RTL with no host in the loop. A 2-layer ReLU network learns XOR on the chip itself, 4 of 4 correct, every node bit-exact from specification to the binary FPGA.

SKY130 — submitted for fabrication through Tiny Tapeout

The design was submitted for fabrication through an open ASIC process: GDS was produced, the gate-level test passed, and precheck passed. No die measurements are claimed.

2.1× / 2.6× — withdrawn, and what stands instead

A float whose exponent is a balanced-ternary number and whose fields are fixed: no regime decode to pay for, and on a ternary fabric the exponent add is native. The accuracy lead over takum once reported here is withdrawn — at equal stored width the comparison ties rather than wins. What survives is measured in hardware, not accuracy: 38× fewer LUT on a full adder against a true base-3 opponent. Range is bounded at ±40 in powers of two; that limit is stated with every result.

Over the air — tri-net, a full ternary network stack

133 formal specifications: GF16 physical layer, BPSK modem on AD9361, ETX mesh routing, ChaCha20-Poly1305 / X25519 crypto. Text and images carried between physically separate boards.

83 formats — a conformance catalogue

Bit-exact test vectors for FP8, BF16, MXFP4 and microscaling formats: a vendor-neutral reference for verifying low-precision arithmetic.

How any of this is checked

Independent model, not a mirror

The reference model is written from the specification, never from the RTL.

Per-stage vectors

Known-answer vectors at every pipeline stage, so a regression points at the stage that broke.

Hardware replay

The same vectors run again on the physical board. Simulation agreement does not prove agreement on the binary FPGA.

Open toolchain

Yosys, nextpnr-xilinx, prjxray, openFPGALoader, iverilog. No proprietary licence stands between a claim here and someone reproducing it.

What these results are not

Entries are entries

A DARPA CLARA submission and an OpenAI Parameter Golf entry are submitted work, not awarded contracts or won prizes.

One device family

Measurements come from a Xilinx Artix-7 and are not multi-corner characterisation.

Training is a primitive

The on-chip training result is proven at small scale: a real network learning on a binary FPGA, not a production training accelerator.

Estimates are labelled

Anything estimated rather than measured is labelled as estimated.

A figure withdrawn

This page previously reported 323 MHz and 41.2 GOPS for the GF16 matmul. Re-checking the RTL on 8 August 2026 showed the block holds no registers, so it has no clock and no frequency can belong to it. Withdrawn rather than explained away.

The papers, the source and a full example report are all public. A claim you cannot verify is just a sentence.