Josh, 5 October 2026 22:05 UTC: if the GPU-server patch does not enable 12 GB cards, research proving and see if there are different methods. docs/analysis/proving-methods.md: (1) the memory anatomy of SP1 6.8.1 from its code (the 20 GB panic in sp1-gpu builder.rs, the maximum-sized trace buffers, the mempool that never releases, the fixed recursion shape), the theoretical floor for the adopted shard (about 10 to 11 GB, approximate); (2) the survey: SP1, RISC Zero, Airbender, ZisK, OpenVM, Pico, Ziren, Stwo, Jolt, Ceno, Nexus, Valida, Powdr, Binius, the 2026 entrants, the sumcheck and GKR family, folding, continuations, distributed proving, AMD and Apple backends, every claim cited; (3) per route the change to the guest, aggregator and node verifier, the cost in agent days, the risk, and whether 12 GB under 60 s is reached; (4) the ranking: re-size SP1's server with S_p as the dial and one server per card (no consensus change), RISC Zero as version 2 for the fallback and Apple, the sumcheck family in five years; (5) the tier consequences and the public line.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
