02Network
Performance and its basis
Numbers without methodology are marketing. Every figure here states whether it is measured, targeted, or not yet known.
Performance, with the basis attached
| Metric | Value | Basis | Measured between, over what | Notes |
|---|---|---|---|---|
| Throughput | TBD TPS | not yet measured | nothing benchmarked yet | Report as a curve across write-account contention, not a single peak figure. Zero-contention transfer benchmarks do not describe real workloads. |
| Soft confirmation | 14 ms | measured | median, submission to the sequencer's signed receipt; 19 ms p90, 22 ms worst. 20 transfers on the gamma devnet chain at 100 ms blocks, 13 November 2026 | Sequencer acknowledgement: a signed receipt at admission, then a block about 100 ms later (96 ms median in the same run). The sender ran on the node's own machine, so the figure leaves out your network: add your own round trip. Revocable until the batch settles: this is not finality. An earlier run on a local test validator, 26 October 2026, measured 16 ms. |
| Ordering to root on Solana | 192 s | measured | median, a transaction's ordering to its block's root committed on Solana devnet; worst 278 s. The gamma chain at 100 ms blocks, a burst of 200 transfers through the public devnet RPC, 13 November 2026 | The Layer 2 side is fast: in the same run a receipt took 14 ms and a block about 100 ms at the median. Settlement is the slow part, and the reasons are ours rather than Solana's. A block larger than about 470 published bytes cannot settle in one Solana transaction: it needs one to open its data batch, one per 900-byte chunk, one to seal and then the root, sent one after another. Blocks settle strictly in order, so a large block delays every block behind it, and the public endpoint rate-limits sends: through a keyed provider the same burst took 64 to 143 seconds. One run of each test on one laptop, and the Solana head is read through a cache that can be 15 to 40 seconds old, so these times may be overstated by up to that much. For comparison, on the earlier beta chain with two-second blocks, a block small enough to travel as one Solana transaction settled in about five seconds, fastest 4.0. Making settlement faster is work in progress. |
| Anchor cadence | 10–15 s | design target | root to next root, streamed batching | How often a state root reaches Solana, once batching streams a frame per L1 slot instead of settling per block (ADR-0006). Today each block's root is settled on its own, strictly in order. Under sustained load on 13 November 2026, with a block every 100 ms, roots reached Solana about once every four seconds, and the backlog then drained at about one a second. Streamed batching is what closes that gap; it is a target, not built. It is also what makes an insured fast lane affordable: the coverage a receipt can carry is sized by anchor latency, so halving the latency halves the capital the promise needs. |
| Fast exit, liquidity-provided | seconds | measured | request to funds on Solana, one provided exit | A provider fills your exit immediately and waits out the window itself. Demonstrated on devnet on 2026-09-04: two withdrawals paid on Solana within seconds of being initiated, against roots days away from finality — and a third refused at the point of use when the provider's float could not cover it, refused rather than queued, with the ordinary withdrawal path unaffected either way. The liquidity was ours and the SOL was devnet SOL, so this measures the mechanism and not a market. A provider's spread is its own price for the risk, and that risk is the window in the row below. |
| Total exit | ~3 days | design target | request to funds, window plus air gap as configured | 48-hour challenge window plus the 24-hour air gap (ADR-0016, final, and ADR-0015) — quoted as the total, because that is what a user waits. The window is 432,000 slots — 48 hours at the 400 ms target — and slots are what the settlement program stores; it is insulated against historic halts because both the window and the dispute deadlines count slots and a halted chain produces none, and a root under an open challenge cannot finalise at all, whatever the clock says. Non-disputed roots are what a fast liquidity bridge can front, and 48 hours cuts a provider's capital lock-up from eight days to three — nobody underwrites that yet, so it clears the way for the market rather than being one. Seven days was the Ethereum convention and was never measured about Solana; the dispute game needs hours, and Solana's four-slot leader rotation removes the multi-hour censorship risk Ethereum's number pays for. The window is fixed when a chain is initialized and there is no setter: the beta devnet chain, initialized 2026-09-05, carries the 48 hours (its settlement program reports 432,000 slots), and the alpha chain it replaced keeps the seven days it was created with. Observed end to end on 2026-09-06: the beta chain's first root finalized 432,000 slots after its commitment — 16 hours on devnet's measured 166 ms slots (the 48 hours is mainnet's 400 ms slot); the parameter is slots by design, so a halt cannot spend it. |
| Fee reduction | TBD × | design target | comparison chain and workload not fixed | Bounded by data-availability cost on Solana. The measured L1 share is about 280 lamports per transaction on modelled sustained traffic; the multiple depends on batch fill, so the cost model is published rather than a single number. |
Benchmark methodology, hardware, and workload mix will be published with the first measured throughput results. Single observations are labelled as such.
01The chains so far
The chains so far
Three devnet chains: the alpha chain the project started on, frozen since; the beta chain that carried the decisions made after it, stopped on 2026-09-12 once its last withdrawals were paid; and the gamma chain that succeeded it on 2026-09-11 and carries them now. Each settled real blocks to Solana devnet, and none is mainnet.
solieum-testnet-gamma
chain 7790 · current
The chain a node built from the main branch runs today, initialized 2026-09-11 as beta's successor. Its account leaves commit data by a page-tree root rather than inline (state commitment 2), which is why it needed a new genesis rather than an upgrade: a node that speaks the new commitment cannot replay a chain written under the old one, and now refuses to by name instead of diverging. Same 48-hour window and 24-hour air gap as beta, the dispute game's enforcer as settlement authority, its own sequencer key with 0.1 SOL bonded under the bond program it shares with beta, and its portal initialized with a guardian on a key that is not the operator's. Its first deposit was credited in block 2; its first withdrawal, 1,900,000 lamports in block 209, was proven against its root once final and paid by the vault on 2026-09-12, an hour after the proof. The vault now holds 795,960 lamports: its rent floor plus the 100,000 that nobody claimed. It cannot go to zero, because the floor is what keeps the account alive. The node now compares what the L2 has committed to withdraw against what the vault can actually pay, and says so. Explorer on port 8895.
solieum-testnet-beta
chain 7789 · stopped
Initialized 2026-09-05, stopped 2026-09-12 at 12:46 IST once its last two withdrawals were paid. It carried the decisions made after alpha: the 48-hour window, the sparse state tree, the enforcer as settlement authority, the fee kept by the sequencer from block 100. It closed at 158 blocks with every root final: two deposits in, 13,000,000 lamports; four withdrawals proven against final roots and paid by the vault on Solana, 14,000,000 lamports, on 2026-09-07, 2026-09-10 and twice on the morning of 2026-09-12; nothing owed. Its bond stands, 100,000,000 lamports and unslashed; an unbond was filed at 13:14 IST the same day and the bond can be withdrawn from 2026-09-19 — stopping a chain does not touch its bond, and filing did not either. It could not run on a node built from the main branch after 2026-09-10, its state commitment predating the change, so it ran its last day on a pinned binary, and its history, 158 blocks archived byte for byte, is readable only with that binary. No explorer: a stopped chain has none.
solieum-testnet-alpha
chain 7788 · stopped
The first devnet chain, initialized 2026-09-02 and no longer running. It carries the seven-day window it was created with, since the parameter has no setter, the sorted state tree, and a dispute authority that is a plain key, so a verdict there enforced nothing. Ten blocks, one credited deposit. Its roots and batches are on Solana devnet permanently, which is what settlement means.
02Reading the table
How to read a performance claim
The table above is deliberately sparse. Here is what each basis label means, and what a number wearing it is worth.
measured
A benchmark ran, and its workload, hardware, and method are published next to the result. You can re-run it. This is the only label that should move money or architecture decisions.
design target
What the system is engineered toward. Honest as an aim, worthless as a promise — targets are what benchmarks exist to confirm or embarrass.
not yet measured
We will not invent a number to fill a table cell. A TBD here is information too: it tells you exactly what has not been proven yet.
03What to wait for
The three numbers that actually matter
Marketing throughput is one number. Using a network is three, and they answer different questions.
- 01
Soft confirmation — "did it go through?"
The sequencer's acknowledgement, in milliseconds. Good enough for a game move or a social action. Revocable until the batch settles, so never enough for an exchange deposit.
- 02
Data posted — "can anyone verify it?"
When the batch is published and independent nodes can re-derive the state. This is when the transaction stops being the operator's word and starts being checkable.
- 03
Finality — "can I build on it irreversibly?"
When the state root is final on Solana and a withdrawal can complete against it. Custody decisions, exchange credits, and settlement all wait for this one.
Why throughput is a curve, not a number
SVM throughput comes from executing non-conflicting transactions in parallel. Transactions that touch the same account — the same order book, the same mint — serialise no matter how many cores exist. So real throughput depends on how much the workload contends with itself: a transfer benchmark with zero contention produces a spectacular number that describes nothing. When Solieum publishes throughput, it will be as a curve across contention rates, with the workload mix stated. Anything less is the crypto equivalent of a car's top speed downhill with a tailwind.
Building something the base layer can't hold?
Tell us the workload. If Solieum is the wrong answer for it, we would rather say so early than have you find out at launch.