Running Core's real consensus code inside a zkVM

The recent benchmark update for block 962,000 provides insightful data regarding the efficiency and timing of transaction processing using a single idle L40S in guest mode 4 (chunk_prove). This specific mode focuses solely on reading the block and its previous outputs without involving any accumulator or bridge, which enables direct measurement of blocks near the blockchain's tip without needing full chain processing.

The detailed breakdown of the processing times across 16 chunks reveals significant variability, with per input processing times ranging from as low as 0.58 seconds to as high as 15.29 seconds. Specifically, chunk 2 exhibited the highest processing time at 7660 seconds, averaging about 15.29 seconds per input. In contrast, both chunks 5 and 6 recorded the fastest processing times, each taking only 292 seconds, translating to approximately 0.58 seconds per input.

This variability highlights different computational demands or efficiencies across the chunks within the same block, indicating areas where optimizations might be targeted. The overall computation time for proving all transactions was 30,385 seconds (approximately 8.44 hours), constituting about 86.8% of the total computation effort. The aggregation of these chunk receipts into a single block proof required an additional 4,606 seconds (about 1.28 hours), making up 13.2% of the total time spent. The complete process for handling this block thus summed up to a total of 34,991 seconds, or approximately 9.72 hours, averaging 4.37 seconds per input.

This comprehensive analysis not only underscores the technical capabilities and limitations of the current system but also sets a critical baseline for future improvements and optimizations in blockchain processing technologies.