AxiomLimit · AI-cluster networking · for a technical reader
Fixed-size memory for reordering AI network data: the lab’s exact wording
The lab’s own sentences and figures for this result, word for word, in its working terms, with its limits in plain words where the lab’s text cannot be reprinted. The plain account is on the result’s page; it says the same things in plain words.
The lab’s plain-English sentence
In a computer simulation of an AI-cluster network, a receiving network card keeps track of … that arrive out of order with a small note of fixed size instead of a buffer that grows with the data in flight
[The lab’s qualifying sentence is not printed here: it states a throughput edge that no file this site serves measures. In the served file, “Meta CTS” names the lab’s simulator model of a second design; no public Meta source states its reorder memory. The “escape sweep” in the limits is a set of adversarial runs against the design.]
Against a published fixed-size design, in the lab’s words
1.965x against faithful bounded-state prior art on the identical grid
The limit to read first, in the lab’s words
netsim simulation, not silicon.
Claim
Bounded descriptor state scales O(1) where the faithful buffering field scales Theta(W), giving a 3,886x robust geomean over the closed-loop grid — and 1.965x against faithful bounded-state prior art on the identical grid.
[Plain gloss of the claim: “faithful” is the lab’s own word for its models of the earlier designs, which nobody outside the lab has checked; the larger ratio is against cards that buffer all out-of-order data, which the result page calls not the fair comparison; the fixed-size scaling is what the simulated grid showed, not a proof; and the smaller ratio is against the lab’s model of STrack alone.]
Limits
Simulation, seeds [1,2,3], generated_utc 2026-06-27. Four sibling instruments are NO-SUM and must never be added: 122,880x, 42,111x, 24,320x, 62,208x. The 62,208x hero divides by exactly the 16 B ratio_floor clamp; the defensible twin is hero_large_message at 39,637x with a 96 B denominator. Three published n for the same geomean reconcile as 960 all rows / 768 wall-payer rows / 756 nonzero. FIX-I open: 122,880x still unlabelled across 17 files. STrack was excluded by name from the escape sweep until 2026-08-19; thresholds were NOT moved to accommodate it.
Against a card that buffers, which grows with the data in flight
giving a 3,886x robust geomean over the closed-loop grid
A larger ratio in the lab’s files, and why it is not used
The 62,208x hero divides by exactly the 16 B ratio_floor clamp; the defensible twin is hero_large_message at 39,637x with a 96 B denominator
More
- The plain account of this result: the problem, what was shown, what it means for a buyer, who we expect would buy, why now, why you can trust the check, and what it does not show yet.
- This result’s file, the source of every sentence above.