EIP-7904 Compute gas cost analysis
No call decision · Fees & gas · EIP status Last Call · created 2025-02-05
An informational EIP recording the conclusion, based on actual measurements of EVM operation and precompile execution times, that no operation needs a gas cost increase to meet the 100 Mgas/s target. The gas schedule does not change.
Why it is needed
Every operation in Ethereum is priced using a fee unit called gas. If an operation is priced cheaply in gas relative to the time it actually takes, that operation becomes a bottleneck in block processing, making it hard to safely raise the block gas limit. An earlier version of this proposal was a Standards Track proposal to raise the gas cost of operations that measurements showed falling short of the target throughput. However, once the client optimizations enabled by Block-Level Access Lists (EIP-7928), parallel disk reads, parallel transaction validation, and execution-free state updates, had all landed in the major execution clients, rerunning the benchmarks completely changed the picture. Operations that had previously looked underpriced now comfortably exceeded the 100 Mgas/s target at their existing gas costs. Since the original justification for an increase had disappeared, this EIP was converted to Informational, preserving the analysis methodology and data while documenting the conclusion that no compute gas cost needs to change.
What changes
Today
- The earlier version was a Standards Track proposal to raise the gas cost of operations measured as underpriced
- Before the EIP-7928 optimizations were reflected, measurements showed some operations and precompiles falling short of target throughput
- These operations were identified as the bottleneck blocking a block gas limit increase
After EIP-7904
- Remeasured on optimized clients, all 18 parameters reviewed meet the 100 Mgas/s target at their current prices
- No change whatsoever is made to the gas schedule
- The benchmarking methodology and data remain on record for future gas rebalancing work
Who is affected
- Users: Since no change is made to the protocol or the gas schedule, fees and user experience do not change.
- Contract and app developers: Since the gas costs of the opcodes and precompiles under review remain unchanged, no action is required in response.
- Client teams: Because the worst-performing client's methodology sets the network-wide baseline, each team's performance becomes the reference for future reanalysis.
- Researcher: The measurement pipeline and data leading from the EEST benchmark through Benchmarkoor to NNLS regression are preserved as the foundation for future repricing discussions.
In depth
Specification: no change to the gas schedule
This EIP specifies no change to Ethereum's gas schedule. Not a single opcode or precompile gas cost is modified. The analysis covers the 18 gas parameters flagged as repricing candidates in previous revisions, and for each parameter it tabulates the current gas cost against the gas cost implied by the worst-performing client at the 100 Mgas/s target. A measured value at or below the current value means the operation already meets the throughput target at its current price. Per-client details and the source of the worst-case values are in the full proposal report (new_gas_proposal.md) attached to the EIP.
Measurement results for the 18 parameters
16 of the 18 were measured as cheaper than the current schedule, and one, PRECOMPILE_BLAKE2F_PER_ROUND, showed no change. Notable drops include PRECOMPILE_ECPAIRING_BASE at 8247 versus 45000 (-82%), PRECOMPILE_ECRECOVER at 841 versus 3000 (-72%), OPCODE_ADDMOD at 3 versus 8 (-62%), and PRECOMPILE_POINT_EVALUATION at 23825 versus 50000 (-52%). The only parameter measured higher than its current value is PRECOMPILE_BLAKE2F_BASE, which is currently unpriced at 0 gas; the measured value of 94 gas is negligible against typical block gas usage and is not a throughput bottleneck. So no operation is underpriced to the point of needing a gas increase.
Target throughput of 100 Mgas/s
The analysis is benchmarked against a performance target of 100 million gas per second (100 Mgas/s). This is the throughput an optimized client is expected to sustain, substantially higher than historical baselines. An operation is judged appropriately priced if its execution speed, measured at the current gas cost, meets or exceeds this rate. Every compute operation and precompile reviewed in this analysis met or exceeded this target under the current gas schedule.
Data collection and the worst-client principle
Similar to the Gas Cost Estimator project, synthetic blocks are built that isolate and intensively execute individual EVM operations for measurement. Blocks are generated by varying the operation count and input parameters, and the tests are collected into the EEST benchmark suite. The Benchmarkoor tool collects total block execution time and operation count, and each block is run multiple times per client to reduce variance. The clients most recently measured are Besu, Erigon, ethrex, Geth, Nethermind, and Reth. Operations whose cost varies with input size, such as KECCAK256, are measured across a range of input sizes. Implementation differences are handled under the general principle of taking the result from the worst-performing client, since that cost cannot be optimized away. Having the slowest client for a given operation define the network-wide cost is the safest and most conservative choice, and the table values follow this standard.
NNLS regression and gas conversion
Non-negative least squares (NNLS) linear regression is used to estimate the time of a single operation out of the block's total execution time. All coefficients are forced to be non-negative so execution time cannot come out negative. Simple operations are estimated as runtime = intercept + slope x operation_count, while operations with a parameter add a param_coef x operation_count x param term to this. The intercept represents baseline overhead including test setup and teardown, and the slope is the per-operation execution time. Models are built independently per client and operation, and only well-fitted models with R² > 0.5 and p-value < 0.05 are used in the conclusions. Gas conversion uses the formula gas = (anchor_rate x runtime_ms) / 1000, with anchor_rate set to 100 Mgas/s. A model-fit report and the core output CSV are attached to the EIP.
Why EIP-7928 changed the conclusion
EIP-7928's block-level access lists enable client optimizations such as parallel disk reads, parallel transaction verification, and execution-free state updates. The latest benchmarks were run on clients that have implemented and stabilized these optimizations, and this is the key reason the analysis concludes repricing is unnecessary. Once the optimizations took effect, compute operations previously flagged as bottlenecks now pass the target throughput at their current gas cost. Note that lowering gas costs is out of scope for this analysis. The analysis's only goal was to find and remove underpriced bottlenecks, and applying the same standard to the latest data leaves no remaining candidates, so the gas schedule is left unchanged.
Limits of the analysis's scope
This analysis deliberately covers only computational complexity, meaning pure CPU execution time, and excludes network-related costs such as state persistence. This keeps the results from being tied to a specific implementation, making them applicable across different EVM clients and simplifying the estimation. It also does not separately address the unique burden of ZK-SNARK proof generation. However, since this analysis recommends no change to gas costs, it has no effect on EVM block proving time either. Related external work cited for context includes the Gas Cost Estimator, which measured seven EVM implementations, EIP-7883, which analyzed ModExp, the Nethermind Gas Benchmarks, measured with a standalone client, and the EVM Memory Analysis, which covered memory costs; the results are consistent with one another.
Status and caveats
This EIP is of the Informational type and is currently in Last Call status, with a deadline of 2026-11-01. Since it recommends no change to the protocol or gas schedule, there is no consensus change and no direct security impact. The analysis itself takes a conservative approach, relying on the worst-performing client per operation and using only regression models with good statistical fit. However, the source text explicitly states that if client performance characteristics change in the future, for example if optimizations are reverted or new operations are introduced, the analysis must be rerun before relying on this conclusion that current compute gas costs are appropriate. The core developer call notes for this material contain no decisions regarding this EIP.