Starknet's Stwo prover cuts peak memory by 85% so phones can prove their own transactions

Divided out, the move from 13.9 GiB to 2.1 GiB is a factor of about 6.6. That is an 85% cut in the most RAM the prover needs at any single moment of a job ^[16][17][15]. The figure applies to transaction proofs ^[10]. Separate benchmarks put Stwo's typical footprint between roughly 10 GiB and more than 20 GiB, scaling with the size of the workload ^[4]. Even the bottom of that range is about 4.8 times the new peak ^[19].

Crypto Briefing noted that a phone proving its owner's own transactions is doing a different job from proving entire blocks, and that the block job is the network's heavy lifting ^[10]. The report does not say what proving costs or who runs the block provers. Starkware's roadmap aims at client-side and decentralized proving at lower cost ^[6]. Client-side proving is what the 2.1 GiB figure delivers. Decentralized proving means spreading the work across many participants instead of a handful of specialized operators ^[7]. Block proofs that climb past 20 GiB need at least 9.5 times the memory of a phone-made transaction proof ^[18].

Verification is unchanged. Anyone can check a finished proof quickly without redoing the work ^[14]. On Starknet, a Layer 2 built on Ethereum ^[11], a proof made on the handset is checked the same way as one handed off to a remote server ^[12].

On this evidence, recursion is the way the phone result could reach the cost side. Starkware is building proofs that verify other proofs, compressing a lot of work into a smaller final package ^[8]. If those proofs can take phone-made transaction proofs as inputs, part of the block workload moves onto handsets. The weaker outcome is that the phone figure stays a benchmark. Crypto Briefing calls the proofs made on thousands of handset models local ZK proofs ^[20], a wider category than Stwo transaction proofs.

I think the trust assumption on the user's side has changed and the cost assumption on the network's side has not. The counter-case is that recursion lets users prove part of each block on their own phones, which would take that share off the operators' machines. A published block-level peak close to the transaction figure would prove this view wrong. So would blocks built from phone-made proofs.