Rendered from docs/reports/TODO.md in yaroslavvb/et-soc1-prototyping by scripts/build-todo-page.py. Part of the ET-SoC-1 measurement reports.

TODO: changes waiting for the next pass

Written 2026-09-26 from a review of every published page and of the repository by eight AI review agents, each finding checked against the page text or the data before it was listed (the request is in ../findings/02-requests.md). The live pages equalled their files on that day (MIRROR.md, "Last check").

Status, 30 September 2026 (13:05 PDT): 134 items: 108 done, 6 superseded by the version-3 results, 20 open (four of them the owner's, in part 0). The review of 26 September listed 110; the lab re-check of 27 September added 7, the major pass of 28–29 September (Q65) 9 more, the sparse parity work (Q66, E59) its 6 next steps and the owner's call on its page, and the third card's runs of 29 September one open result.

Status, 14 October 2026 (week 42): the item counts are unchanged — 134 items: 108 done, 6 superseded by the version-3 results, 20 open, four of them the owner's in part 0 — but part 0's lab-fixes item has moved. Six of its eleven numbered fixes are done: U16, U17, U18, U20, U24 and U27, each checked read-only on all three hosts on 14 October 2026 (the notes sit with the items). U19, U22, U23, U25 and U26 were not checked and stay open. Both cards that were out of service are back: aifoundry1's card 0 since 2 October 2026, when its broken fan was replaced, and aifoundry2's since 9 October 2026, on a two-fan 92 mm bracket fitted at the card. The live page had last been built from this file on 30 September; it is rebuilt with this change.

0. For the owner

Proposed: - Replace §4 with "About 400 GB of the 452 GB pool is user data in home directories; the per-user breakdown went to the lab admin." - Write "a user's et-platform checkout" for the named paths. - Remove the transcript from the tree.

Rewriting git history, and whether the two pages stay public, are the owner's decisions. Nothing was changed.

Done 27 Sep, by the owner's decision: the two aifoundry1 pages and the disk transcript stay public for now; the proposed changes were not made.

A. Page changes (rebuild and deploy)

Charts: done on 26–27 September (commits 6f207ce to 974d120, then the merge 0cfc742). Every "Chart"/"Charts" item below is built, except two: heat per mm's optional per-step increments and a chart of the power page's SP pass per card (V3-TEL's values are on that page as text). On-chip communication's grain calculator was added in the merge. Many fix and cut items landed in the merge too; those are ticked below.

Across pages

Hub · sources/limits-of-observability.*

Energy manual · sources/energy-manual.* (markdown through tools/ettelem/render_*.py)

Heat per millimetre · sources/heat-per-mm.*

The DVFS loop and its leakage · sources/dvfs-leakage.*

Power and temperature telemetry · sources/power-temperature.*

The Horace experiment · sources/horace-experiment.*

Why is it low power? · sources/why-low-power.*

One hot line stops a shire · sources/hot-line.*

Hand it to the next shire · sources/on-chip-relay.*

On-chip communication · 2026-09-18-et-soc1-on-chip-communication.html

L2 mainline starvation (pointer page) · 2026-09-22-et-soc1-l2-mainline-starvation.html

Anatomy of a memory access · workloads/memprobe/report_template.html

Memory hierarchy · 2026-09-18-et-soc1-memory-hierarchy.html

Ridge points · 2026-09-18-et-soc1-ridge-points.html

Matmul efficiency · 2026-09-18-et-soc1-matmul-efficiency.html

Sparse compute · 2026-09-18-et-soc1-sparsity.html

Test drive · docs/report/index.html

Spatial temperature brief · 2026-09-22-et-soc1-spatial-temperature-brief.html

Influence functions · sources/influence-on-et.*, data/2026-09-25-influence-on-et/make_analysis.py

The aifoundry1 pages (standalone HTML)

The ET-SoC-1, interactively (the chip diagram) · sources/chip-diagram.*

B. Repository-only work not done in this pass

After the campaign (every block hashes tools/claims-v3/)

Left: all five. - ☑ With the results: - Landed: E35–E49 registered and "Cite as" widened (03-experiments.md); the requests of 25–27 September numbered (Q44–Q57); a results series in CLAIM_SERIES; AGENT.md and getting-started updated. - Left: rebuild aifoundry1's build/<workload> directories with the g3log fix. - Done 28 Sep: the rebuild (the next section's first item).

Our tools, after the heat-placement work (from the lab re-check of 27 September)

The heat code is frozen until its validation on aifoundry1's card 1 ends (about 08:00 PDT, 28 September). Then:

The knowledge base against the version-3 pages (the page holds)

docs/findings/ predates the version-3 first pass (0399048). This pass fixed the summaries in its commit and flagged the rest at the top of ../findings/README.md.

Records and commands

Concision in the repository (about 4,000 words)

C. Card work and its records, after the major pass of 28–29 September (Q65)

D. Sparse parity (Q66, E59): next steps

E59 solved noisy sparse parity on aifoundry3's card on 29 September: L1 (512, 4, 0.3, 448) in 0.131 s, L2 (512, 4, 0.4, 1,850) in 0.323 s and (256, 5, 0.4, 1,925) in 1.52 s on 1,024 minions, 1.1–1.6× six tuned AVX-512 threads (../findings/03-experiments.md, E59). What limits it (workloads/sparseparity/data/2026-09-29-aifoundry3-card-m4/README.md and -m5-energy/README.md): hart 1's row generation, the shire's bandwidth for streamed A (about 512 cycles per op at 32 minions per shire), and the epilogue (1,606 cycles per output tile with the tensor unit idle). The code, the builds and the card rules are in workloads/sparseparity/README.md; every card step is one locked timeout 10 process.