trisha roadmap

trisha closes the loop: source → compile → run → prove → verify → deploy.

status

what's real

full prover pipeline on GPU. 7 compute shaders dispatched to Metal/Vulkan/DX12 via wgpu. kernel fusion eliminates per-layer CPU↔GPU sync. VRAM budget auto-detected from adapter limits; oversized buffers gracefully fall back to CPU. dependency patching injects GPU hooks into triton-vm without maintaining a fork. 4.3× proving speedup on small programs.

neptune HardforkBeta CPU mining: 0.33 MH/s on M4 Max (rayon + triton-vm Tip5). honeycrisp GPU mining: 1.1 MH/s (39 Tip5 permutations per nonce; Metal MSL kernel; compute-bound at ~65% GPU utilization — practical ceiling for this algorithm on M4 Max). GPU path activated via --features gpu; trisha mine --neptune --bench-gpu 10 benchmarks. CPU+GPU combined: ~1.43 MH/s. wgpu Tip5 + Mine.wgsl shaders exist and compile but not wired into neptune_mine.rs.

what's scaffold

deploy (prints metadata, no blockchain interaction). honeycrisp Phase 2 (NEON Tip5 via acpu) and Phase 3 (unimem zero-copy IOSurface).

proposals

confidence milestone

# cpu pipeline (passing)
trisha run /tmp/hello.tri
trisha prove /tmp/hello.tri --output /tmp/p.toml
trisha verify /tmp/p.toml

# batch (passing)
trisha run batch /tmp/a.tri /tmp/b.tri
trisha prove batch /tmp/a.tri /tmp/b.tri --output proofs/
trisha verify batch proofs/a.proof.toml proofs/b.proof.toml

# GPU (passing)
trisha prove /tmp/hello.tri  # dispatches NTT/Merkle/FRI/GEMV/Hash on GPU

# mining (passing)
trisha mine /tmp/hello.tri --difficulty 1000000000000000000

# not yet passing
trisha deploy /tmp/hello.tri --state testnet
trisha merge proof1.toml proof2.toml --output merged.proof.toml