The patent isn't WebGPU. The patent is the routing decision made on every call. Below, the same sortAsync function gets different backends as you change the input. Drag the slider or pick a scenario - the scorer publishes its workload profile, its seven weighted factors, and the final tier. Then press Race the lanes to actually run the workload three ways.
sortAsync(arr) - is routed to a different backend depending on the array's shape. Vanilla WebGPU doesn't make that choice; it pays the upload cost on every call. Hadron's scorer publishes its seven factors and the named reason for every dispatch, so the decision is auditable. This is what the patent filings protect.
Five UK patent applications, all assigned to Ayoob AI Ltd, covering the cross-platform dispatch layer. Each Play button runs a short sequence that shows the exact mechanism the patent protects.
Same function, eight possible paths, picked per call.
A stream of six differently-shaped sort calls. Each is routed to a different backend based on its workload profile - decision published per call.
Hadron refuses GPU when Float32 would lose accuracy.
Two matmuls, identical shape, different value range. Run 1 fits Float32; Hadron goes GPU. Run 2 overflows Float32; Hadron refuses and falls back to Float64 CPU.
Each operator scored independently; GPU intermediates stay GPU-resident.
A 1M-row pipeline: filter → groupBy → sort. Each box gets its own tier. When two consecutive operators are GPU, the intermediate buffer stays on-device (the core dispatch routing).
Pre-filter on GPU (cheap), verify on CPU (precise).
2M chars of synthetic text. Phase 1 runs a pre-filter pass on GPU; rules out 95% of the corpus in one pass. Phase 2 verifies the remaining 5% on CPU with a CPU literal-search path.
Six named structural reasons. Each refusal cites a different one.
Four search patterns submitted in sequence. The literal goes GPU. The other three are refused with named architectural reasons - each refusal cites a distinct hardware constraint surfaced via the dispatch info.
For deep-dive demos. 34 isolated tests grouped by library / patent. Use these for technical Q&A.
Multi-factor dispatch routing for sort workloads. Profiles each array and routes to the most suitable tier. Sync sort stays CPU; sortAsync adapts.
matmul / FFT / conv2d / solve auto-dispatch across tiers. Precision-sufficiency check before GPU; pipeline fusion keeps intermediates on-device.
Each operator in a pipeline is scored independently. Consecutive on-device operators retain their intermediate; mixed pipelines pick the best tier per step.
Substring queries dispatch to the right tier per shape. Regex / wildcard / fuzzy / long-pattern / Unicode-case-fold / high-density workloads are categorically refused on GPU with a named reason. Multiple distinct refusal categories, each surfaced via the dispatch info for audit.