The first GPU router in which the parallel routing itself produces legal results, independently verified.
Legal FPGA routes across four public benchmark designs, from a small SoC to a Titan-class miner, each checked by an independent verifier.
Measured results
Four public FPGA benchmark designs. Every figure below is measured on the routed output.
| Design | Nets | Result | Detail |
|---|---|---|---|
| picosoc | 4,055 | legal in ~0.5 s | 2.4x faster than VPR |
| neuron | 50,835 | legal in ~12 s | 1.25x faster than VPR at like-for-like width |
| sparcT2 Titan class | 182,090 | legal | verified 3 of 3 runs |
| bitcoin_miner Titan class | 799,403 13.7M-node graph | legal in ~259 s | single 96 GB workstation GPU |
The honest split
Faster than the reference on small and medium designs, both fully legal; at the largest scales the demonstrated property is verified legality on a single GPU, which no published GPU router has.
Verification
Every routed result is checked two ways: an independent occupancy recount and a full connectivity check. Every run is independently verified legal; a run that does not reach a legal route reports it and exits, so a non-legal result can never pass as legal.
On sparcT2, the router reached a legal result in 3 of 3 runs.
Run it yourself
The reproduction kit routes all four public benchmark designs on a Blackwell-class GPU and re-checks each result with the same independent verifier. The designs regenerate from the public VTR/Titan benchmarks, so you build the instances yourself, nothing to take on trust. Available to evaluators on request.
Watch a routing run
LiveWhitepaper
Full method and results
The complete technical brief: measured results across four designs, test conditions, the independent verification method, and the honest width standing against VPR.
Technical brief (PDF)Patent status
Filed
U.S. Patent Application No. 19/731,882, filed July 2026.
Talk to the team
For technical detail, benchmarking on your own designs, or licensing, get in touch.
Contact Voxell