Prototype Reduces Ethereum Blob‑Recovery Computing Load by Up to 18×

A prototype that distributes blob‑recovery duties among nodes reported a dramatic reduction in the estimated computing effort required to rebuild missing data. The study, authored by researcher Csaba Kiraly, suggests a smaller‑scale change could lessen the workload without adopting the full RowDAS networking proposal.

How the reduced design works

Ethereum’s PeerDAS system lets nodes verify that rollup data blobs are available by downloading only a portion of each blob. High‑custody nodes store at least 64 of the 128 data columns, while supernodes store all 128 columns. In the reduced design, specific high‑custody nodes are assigned particular blobs to recover first, allowing other nodes to receive the reconstructed data instead of duplicating the effort.

Simulation results

Across 1,000‑node simulations, the reduced design achieved an 11–18× reduction in total CPU‑seconds needed for reconstruction:

• With four blobs, 10% supernodes and no columns withheld, network‑wide CPU cost fell from 48.6 seconds (PeerDAS) to 2.75 seconds.
• With 20% supernodes, the cost dropped from 91 seconds to 6.6 seconds.

The figures represent accumulated work across the network, not elapsed time, and are based on a measured 162 ms cost per blob recovery on a Ryzen 9 8945HS processor.

Limitations and next steps

The reduced design retains reliance on high‑custody nodes and does not add the additional resilience offered by the full RowDAS proposal (EIP‑8371), which would introduce row‑channel networking for collective reconstruction. The current measurements are limited to simulated environments using real cryptography; larger‑scale simulations and real‑network testing are still required. The draft EIP‑8371 has not yet incorporated the duty‑assignment split.

Implications

While the immediate benefit is a narrower reduction in processor work for blob recovery, the approach could serve as a stepping stone toward broader resilience improvements once the full row‑networking layer is implemented.