mirror of
https://github.com/lux-apps/papers.git
synced 2026-10-30 18:15:21 +00:00
zeekay
de41e5d072
Retract the fabricated '581M/597ns / 100M+/s / 434M GPU / planet-scale' headline (a time.Sleep sim + x40 fudge that measured nothing) across the corpus. Flagship lux-lightspeed-dex rewritten (decomplected d-chain engine / dexvm atomic proxy / Go wrapper / 0x9010 CLOB facade; measured GB10 104-126M & M4 Max ~250M fills/s byte-parity-gated; Knuth-D division lever; Block-STM settlement; P3Q->starkfri rollup to Z-Chain). Technical papers (precompile-engineering, gpu-crypto-stack, evmgpu-benchmark, p3q-corona, smart-order-routing, fhe-mpc-hybrid) carry the measured impl + cross-ref the flagship. >1B/s framed honestly as N-node aggregation, never single-GPU. All compile.
26 lines
1.7 KiB
TeX
26 lines
1.7 KiB
TeX
\section{Conclusion}
|
|
\label{sec:conclusion}
|
|
|
|
Lux Lightspeed DEX makes the matching engine a chain. The D-Chain holds matching and state;
|
|
a stateless atomic proxy moves value; a Go wrapper is the client; an inert-by-default EVM
|
|
precompile is a CLOB facade. The open-source node validates the public chains and excludes
|
|
the D-Chain, so the open network never depends on the proprietary engine. Determinism is
|
|
enforced by running the matcher in \texttt{Block.Verify} and carrying fills in the block
|
|
bytes, so accepted state is a pure function of the block.
|
|
|
|
The performance is measured, not asserted. A byte-parity-gated GPU CLOB matcher sustains
|
|
104--126\,M fills/s on an NVIDIA GB10 and $\sim$250\,M on an Apple M4 Max, with the dominant
|
|
lever an algorithmic one---replacing a 512-iteration bit-serial division with Knuth
|
|
Algorithm~D. Settlement runs market-sharded Block-STM with a byte-identical CPU/GPU commit
|
|
order. The rollup to the Z-Chain verifies one proof per $K$ trades---P3Q at 4{,}714
|
|
verifies/s, \texttt{starkfri} at 986 for a toy AIR---and clears $10^{9}$ settled-tx/s on the
|
|
finality axis at fleet scale.
|
|
|
|
We are explicit about the ceilings: a single GPU does not match $10^{9}$ byte-exact
|
|
\texttt{uint256} orders per second, and that figure is reached only by aggregating measured
|
|
single-node rates across a fleet or by a cheaper representation not yet built. The earlier
|
|
``581\,M orders/sec at 597\,ns'' and related figures are fabricated and retracted. The numbers
|
|
here are measured on stated hardware under a parity gate; everything else is labeled an
|
|
aggregation, a model, or a projection. That is the contribution: a decomplected DEX with
|
|
honest, reproducible performance.
|