ET-SoC-1 · Brief · 22 September 2026 · superseded

L2 mainline starvation: superseded by One hot line stops a shire

22 September 2026 · no card: an argument from a reported result, the shire-cache spec and the errata · part of the ET-SoC-1 measurement reports

This brief argued from a result Ivan reported in the AI Foundry Discord (shire 0 got 6% of its fair share of a contended global atomic) and from the shire-cache spec, by which requests arriving from other shires outrank the owner shire's own, that the owner of a hot line is starved. It called the l3_yield priority the lever, advised homing hot structures in a shire that does no compute, and warned that a 2D TensorSend array can hang a hart. The reproduction that afternoon, on two cards, kept the aggregate (60 million atomics a second) and the even split among the other shires, but not the 6%: the owner's share of the atomic is 1.004. What stops, with the clock steady at 600 MHz, is the owner shire's other memory traffic, to 0.01–0.05% of normal in a 10 ms window, once more than 21 remote requesters saturate its bank (22 stopped it, 21 did not, on all three cards checked). Two errata describe the mechanism and limit what l3_yield could do. The placement advice holds, and so does the TensorSend warning for cells that take two partners at once, which rests on the RTL and on one hang seen on aifoundry2.

Checked on three cards (26 September 2026): the two claims this brief quotes from One hot line stops a shire were re-measured under a pre-registered plan, in three passes on each of aifoundry2, aifoundry3 and aifoundry1 card 1. They hold, with sharper numbers: the owner's traffic stops at 0.01–0.05% of normal in every configuration that stops it (the low end, the DRAM-homed line, now repeated on every card), the edge is exactly 22 remote requesters, where the hot line's inequality puts it, and pacing gives the owner back 40–41% just above the 9,704 cycles the rule asks for and 55–57% at 10,000. Of 2 claims tested here: 2 held, as the hub’s scoreboard counts them. Record: docs/reports/data/2026-09-25-claims-v3.

Related reports: where its points now live

• One hot line stops a shire — the measurement: the atomic is fair, the owner shire's own loads stop (at a steady 600 MHz), and a bank rule that predicts when (a step-through drawing of the owner shire's bank). Its advice is in What to do instead (section 7): make the owner a shire that does no compute, not the master shire; pace the requesters; spread instead of sharing; and read the victim's bank counters from user mode. Section 8 says why l3_yield was left alone.
• On-chip communication, “A trap: one ready flag per minion” — why a 2D TensorSend array whose cells take readies from two partners at once can hang a hart until the chip is reset (read from the RTL; one such hang was seen, on aifoundry2), and the layouts that are safe: 1D rings and the hardware trees, and, by the RTL rule (not run), 2D in barrier-separated phases.

The brief as corrected on 24 September is kept in the repository, at commit 49dd0c8. Versions: 22 September, first published, before the reproduction; 24 September, corrected in place; 25 September, replaced by this pointer (its points moved to One hot line stops a shire), and the quoted results given their home pages' qualifiers (version 3); 26 September, the quoted results re-measured on three cards; 28 September, the safe 2D layout marked as the RTL's rule, never run, and the note's counts given both ways.

← All ET-SoC-1 measurement reports