igneum-network/igneum
1
0
Fork 0
Code Issues Pull requests Releases Activity

Counter ASIC 3.0 status: the two claim drafts for Josh with the rows that bound them (main's steering)

Browse source
This commit is contained in:
igneum-josh 2026-10-06 08:35:09 +01:00
parent 870363c0c7
commit 3ffa208769
1 changed files with 15 additions and 1 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

16
docs/plans/counter-asic-3-status.md
View file

@ -67,7 +67,21 @@ What moves the f = 1 rows (chip-model-v3.md section 5.7): not the dataset size (
## 6. Decisions for Josh
1. The public claim. "Under 2x" is true of the recompute chip per chip at the op budget and false of the stored-dataset chip per joule (item 1). Re-word to the measured fact (the honest card runs at 1/128 of its dependent-read ceiling; a chip must out-read it per watt) until the item 8 rows land, or keep the claim with "per chip, against the recompute chip" stated. Nothing changes on the site until Josh's word.
1. The public claim. "Under 2x" is true of the recompute chip per chip at the op budget and false of the stored-dataset chip per joule (item 1). Nothing on the site or in the litepaper changes from this run; the two drafts below are for Josh's decision, with the rows that bound them.
| Row | Per chip (rate) | Per joule | Source |
|---|---|---|---|
| On-die recompute chip, class v3 (f = 0) | 0.92x with the 3x factor (0.31x bare) | 1.86x (1.3x to 2.4x over the on-die read energy) | chip-model-v3.md sections 2 and 5.4, model |
| Stored-dataset memory-controller chip, GDDR7 (f = 1) | 1.22x | 5.1x | chip-model-v3.md 5.4, model |
| Stored-dataset chip, HBM3 one stack / eight stacks | 0.61x / 4.9x | 7.5x / 9.2x | chip-model-v3.md 5.4, model |
| Ethash precedent, the same chip class: Linzhi Phoenix 2020, Antminer E9 2022, Jasminer X4 2021 | | 2.1x / 2.9x / 4.8x per joule | asic-resistance-history.md rows 3 and 4 |
| With the latency shadow filled (item 8, model until measured) | 1.22x | 1.85x at N = 330,000 and a chip core equal to the GPU's ALU; 2.5x at 1.5x worse | chip-model-v3.md 5.7 |
Draft (a), scoped: "The strongest recompute chip we can price, holding the whole 256 MiB cache on-die, reaches under 1x per chip against an RTX 5090. A memory-controller chip that stores the whole dataset reaches 1.2x per chip and, in our model, 5x to 9x per joule; the Ethash chips of this class reached 2.1x to 4.8x. The lever against it, program work in the latency shadow, is being measured (Counter ASIC 3.0 item 8)."
Draft (b), the measured fact only: "An RTX 5090 mines this hash at 136 MH/s and 17.5 billion dependent 4-byte reads a second, 82 percent of its memory's random-read ceiling, with its integer units at 0.15 percent of their budget. A chip beats it only by reading per watt what a 512-bit GDDR7 board reads, and the gap is the card's own idle logic."
Either replaces "under 2x" on the site and in the litepaper once Josh chooses; until then the claim stays as it is and this file records that it is not safe as worded.
## 7. Unverified and owed