Hi all,
I’m Rohan Indurkar. I’ve built Split Optimizer, a routing engine on PropellerHeads’ open-source Fynd and Tycho, with
a proprietary optimisation algorithm for splitting large orders across pools and paths. Before going further
I wanted outside eyes on how it compares, so here is a benchmark anyone can check.
What was measured
- 3,429 sell orders whose pair and size come from executed CoW Swap trades, from about $13k to $12 million (median $72k).
- On 18 Nov 2026 each order was quoted live by three DEX aggregators, each with more than $100B in all-time volume
on DefiLlama. Split Optimizer quoted the same order on the exact on-chain state each aggregator priced. - Same liquidity on both sides: only on-chain venue types Split Optimizer covers (Uniswap V2/V3/V4, Curve, Balancer
V2/V3, Fluid, Ekubo, ERC-4626 and others). RFQ and private market makers excluded, aggregator fees added back.
Results (after gas, size-weighted; positive = more output for the user)
| Orders | Split Optimizer ahead / behind | Size-weighted | |
|---|---|---|---|
| Aggregator A | 3,303 | 2,075 / 169 | +4.8 bps |
| Aggregator B | 3,289 | 1,838 / 249 | +1.0 bps |
| Aggregator C (100+ ETH orders only) | 656 | 41 / 395 | +0.2 bps |
The gap grows with size: on orders of 300 ETH or more it is +5.4 bps against A and +1.2 bps against B. Against C,
Split Optimizer is slightly behind on most orders but ahead in total, because it wins on the largest ones.
Meant to run in parallel. The optimisation is heuristic, so on some orders it returns less than the aggregator (the “behind” counts above). It is meant to run alongside an existing router, keeping the better quote per order, so those orders lose nothing. The heuristics will be tuned on more order data and improved in later versions.
Run both, keep the better quote (after gas, bps of the whole volume)
| Gain | Extra output | |
|---|---|---|
| Aggregator A | +5.3 bps | $433k |
| Aggregator B | +1.8 bps | $150k |
| Aggregator C (100+ ETH orders only) | +1.1 bps | $67k |
Median time per quote is about 120 ms (p90 about 150 ms), measured on a single 16-vCPU Google Cloud VM (c4d-highcpu-16). This is the first version and it will keep improving: latency can be reduced further with more cores and tuning, and throughput can be raised as needed by running more instances, since each one is stateless.
How to check it
- Every order links to its source trade on Etherscan, the block it was priced on, and every pool in Split
Optimizer’s route: [Split Optimizer Benchmark] - The results and each aggregator’s raw API responses are committed by this SHA-256, published here today:
7bbd402d13bdb43a30b478ca0a97a5f59a9598aa633689f8c31593e245f08f2c - Each aggregator can get its own raw responses (with request ids) to verify against its logs.
- I can set up a test endpoint so solvers can compare live auctions themselves.
I’m looking for feedback on the method, and I’d like to talk to solver teams who might use the router. Reach me
at rohanindurkar92@gmail.com or on Telegram @mrr0b0t2004.