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

Class v6 10.0t: the 32-lane 18-family core's synthesis row (k 0.39 same-node, a sixth under the headline placed; the same-node bracket 1.5x to 2.1x, inside the served sentence)

Browse source 
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
This commit is contained in:
igneum-labs 2026-10-08 17:02:26 +00:00
parent 1847d41bf0
commit 5656f6eeef
1 changed files with 1 additions and 1 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

2
docs/design/class-v6-rotating-family.md
View file

@ -603,7 +603,7 @@ The placed 18-family core: 9.36 pJ per lane-op on the class v4 draw at ASAP7 (8.
| The stored-half hybrid, at the hash lane's reconciled mean hit rate of 0.581 | **1.93x** (1.65x to 2.12x); the p98 program 2.09x | 2.40x | 2.80 (5.7x); the p98 program 1.88 | | modelled on the placed core, measured hit rate |
| The N2 SRAM die | 2.3x | 3.1x | 2.94 (5.4x); 2.14 at the smaller ticket | | modelled on the placed core |
**The honest same-node bracket is 1.5x to 1.9x per joule and 3x to 6x per dollar; a node ahead 1.8x to 2.4x.** Re-scored on lane 3's conditions at these energies and dollars (this lane's reading on lane 3's 12a and 8a rows, approximate; lane 3's own re-run owed): the board passes all seven as before (its ratios fall again, its dollars do not move). The hybrid: (b) MOVES TO PASS (USD 2.80 per MH/s is 0.34 of the 5080 entrant's hardware, above the quarter line, where 1.88 was 0.23); (a) passes more clearly (its all-in rises with the dearer hardware and the smaller per-joule gap, toward parity with the best owner at 0.12); (c) still fails (a third of the chain USD 7 to 51 M against USD 12 to 400 M a year, 2x to 8x under it); (e) passes (1.93x); (f) still fails on flat and shrinking paths. The die: (b) MOVES TO PASS (USD 2.94 is 0.36 of the entrant's, where 0.8 was 0.07); (a) still fails (about 2x); (c) still fails (USD 7 to 54 M against 12 to 400 M); (e) passes node-for-node (2.3x) and sits on the line a node ahead (3.1x); (f) still fails. **No verdict moves: the success statement holds for the pure board; the hybrid coexists only in a growing chain at cheap GPU electricity; the die fails once sunk on (a), (c) and (f). What moved is the dollar condition (b), which the adversary lane's own corrected tickets (USD 2.80 and 2.94 per MH/s, against 1.88 and 0.8) now pass for both; the fleet-cost condition (c) and the margin condition (f) are what hold the verdicts.** The served energy sentence (10.0h) reads from this table: 1.5x to 3.1x as complete machines, 1.5x to 2.3x on the GPU's own node. Pending with clocks: the 32-lane full core synthesised (19:30 UK) and placed (21:00).
**The honest same-node bracket is 1.5x to 1.9x per joule and 3x to 6x per dollar; a node ahead 1.8x to 2.4x.** The 32-lane 18-family core's synthesis row (the adversary lane, 18:02 UK, document 0be29c224; ASAP7, a gate-level VCD, the SRAM term modelled, node factors claimed): 393,036 cells and six macros, 5.73 pJ per lane-op on the class v4 draw (5.09 to 7.20), 4.01 at N5, 2.89 at N3, k 0.39 same-node and 0.28 a node ahead at the 5090's lock, 13 percent under the 8-lane full core; the bank's cost at 32 lanes 8 percent of energy and 50 percent of cells. With the measured 8-lane placement factor (1.42) about 5.7 pJ at N5 placed, k 0.55 same-node: the adversary's cheapest placed row sits a sixth under the headline, so the GDDR7 board reads about 1.6x same-node (1.4x to 1.9x) and 2.0x a node ahead, the stored-half hybrid about 2.1x and 2.6x, USD per MH/s unchanged (the lane count is a tenth of the board's capex); **the same-node honest bracket across the lane-count choice is 1.5x to 2.1x per joule**, inside the served sentence's 1.5x to 2.3x on the GPU's own node. The placed 32-lane row lands by 21:00 UK or the next morning (the floorplan with six macros running since 16:26). Re-scored on lane 3's conditions at these energies and dollars (this lane's reading on lane 3's 12a and 8a rows, approximate; lane 3's own re-run owed): the board passes all seven as before (its ratios fall again, its dollars do not move). The hybrid: (b) MOVES TO PASS (USD 2.80 per MH/s is 0.34 of the 5080 entrant's hardware, above the quarter line, where 1.88 was 0.23); (a) passes more clearly (its all-in rises with the dearer hardware and the smaller per-joule gap, toward parity with the best owner at 0.12); (c) still fails (a third of the chain USD 7 to 51 M against USD 12 to 400 M a year, 2x to 8x under it); (e) passes (1.93x); (f) still fails on flat and shrinking paths. The die: (b) MOVES TO PASS (USD 2.94 is 0.36 of the entrant's, where 0.8 was 0.07); (a) still fails (about 2x); (c) still fails (USD 7 to 54 M against 12 to 400 M); (e) passes node-for-node (2.3x) and sits on the line a node ahead (3.1x); (f) still fails. **No verdict moves: the success statement holds for the pure board; the hybrid coexists only in a growing chain at cheap GPU electricity; the die fails once sunk on (a), (c) and (f). What moved is the dollar condition (b), which the adversary lane's own corrected tickets (USD 2.80 and 2.94 per MH/s, against 1.88 and 0.8) now pass for both; the fleet-cost condition (c) and the margin condition (f) are what hold the verdicts.** The served energy sentence (10.0h) reads from this table: 1.5x to 3.1x as complete machines, 1.5x to 2.3x on the GPU's own node. Pending with clocks: the 32-lane full core synthesised (19:30 UK) and placed (21:00).
**Lane 3's re-run at the final energies and corrected tickets (the model's 12b, class-v6-floor-sram 954db053, 17:15 UK, with the workbook rows) replaces this lane's reading above, and a verdict moves: the die's, and only at the adversary lane's ticket.**