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

Class v6 10.5: the 5090 W = 16 row measured (the hinted sparse form holds the rate for +88 W, energy +25 percent; the verdict holds on every card)

Browse source 
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
This commit is contained in:
igneum-labs 2026-10-08 12:38:48 +00:00
parent 9e603e3af6
commit 56276ff039
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/design/class-v6-rotating-family.md
View file

@ -294,7 +294,7 @@ The table, filled with the defaults, each cell replaced as its lane's row lands
| Honest tier (measured joules per hash, class v4) | GDDR7 board (0.466) | HBM3 one stack (0.321) | SRAM die at W = 4 (0.051) | With SM-sparse | W = 16 variant |
|---|---|---|---|---|---|
| RTX 5090 stock (3.36; class v3 2.26) | 3.3x default / 5.8x floor | 3.9x / 7.8x | 5.6x / 21x | 1.3 to 3.9 percent at best, MEASURED (10.1, 14:10 UK: four rented 5090s, the ceiling held to 43 of 170 SMs; the 4090 saves 4 W at 16 SMs, the H100 1.4 percent); the residual is the clock domain, which only the lock takes off; the fraction a chip cannot strip is 16 to 19 percent of the 5090's stock joules, 25 percent at the lock | not applied: KILLED on measured rows (10.5, 13:47 UK): the hinted form holds the rate on Ada but the card pays the second sector at 8 to 9 nJ, energy per hash +34 percent (4090), +33 (H100), the chip +33 (GDDR7); the edge does not move; the 5090 stock row and the PC 1 knee row owed as amendments |
| RTX 5090 stock (3.36; class v3 2.26) | 3.3x default / 5.8x floor | 3.9x / 7.8x | 5.6x / 21x | 1.3 to 3.9 percent at best, MEASURED (10.1, 14:10 UK: four rented 5090s, the ceiling held to 43 of 170 SMs; the 4090 saves 4 W at 16 SMs, the H100 1.4 percent); the residual is the clock domain, which only the lock takes off; the fraction a chip cannot strip is 16 to 19 percent of the 5090's stock joules, 25 percent at the lock | not applied: KILLED on measured rows (10.5, 13:47 UK): the hinted form holds the rate on Ada but the card pays the second sector at 8 to 9 nJ, energy per hash +34 percent (4090), +33 (H100), the chip +33 (GDDR7); the edge does not move; the 5090 measured too (13:40 UK: the hinted sparse form holds the rate at +2.3 percent for +88 W, energy +25 percent, the GDDR7 edge 4.7x to 4.4x at zero shadow and unchanged with the shadow); the PC 1 knee row a formality |
| RTX 5090 at the 1,300 MHz lock, the record's operating point (2.33 = 1.67 + 0.65) | 2.9x / 4.0x | 3.6x / 5.4x | 3.9x absolute (6.1x in the record's convention, marked) / 14x | no change (the lock pass on PC 1 running; the knee row by 15:00) | not applied (same default) |
| Apple M5 Max (1.40; class v3 0.78; the GPU and DRAM channels) | 1.8x / 2.4x | 2.2x / 3.2x | 3.9x / 8.7x | not applicable (no SM lever on Apple) | not applicable (the M5 Max pays 0 at 64 bytes, measured; the chip +33 percent) |
| RTX 5080 at its 1,100 MHz lock, the honest NVIDIA floor (2.06 class v4 measured: 71.20 MH/s at 146.6 W; class v5 2.10; floor lane 4, 10.4) | 2.6x / 3.6x | 3.0x / 4.8x | 3.6x / 13x | no change | not applied |
@ -492,6 +492,20 @@ Nothing reads k above 1 on a measured GPU figure; the second sector's first read
The reading against the identity: the exported form's -47 percent is an occupancy and request-shape effect, as the probe said (the hinted one-request form holds the 4-byte rate on Ada at 98 percent; the sparse shapes recover half the loss on Hopper and Ampere), but the card pays the second sector at 8 to 9 nJ (the 4090: +71 W over 8.3 G reads a second), not the 1.5 nJ modelled: a second sector through the card's L2 and crossbar costs a whole dependent read, so the card's energy per hash rises 34 percent (4090) and 33 percent (H100 best form) while the modelled chip rises 33 percent (GDDR7) and 50 percent (HBM3). At stock the zero-shadow edge does not move on Ada (+34 against +33) and moves 0.9x in the chip's favour against HBM3 on Hopper; the sector's k is 1.15 nJ over 8.6, about 0.13, under the ALU shadow's band. **KILL at stock on measured rows; W = 16 stays out of class v6.** The one row that could still save it is the knee (the card's fixed share is smaller there and its fabric runs at a lower voltage), which a rented pod cannot read (-lgc refused) and which the hash lane's PC 1 lock row answers if anyone still wants it; the 5090 stock row lands when a secure-cloud pod answers (the community 5090 host failed cuInit on three pods) and re-reads the exported form's -47 percent beside the hinted form. Pods and logs in the lane's file; spend about USD 6.
The 5090 row (13:40 UK; a secure-cloud rented 5090 at stock, driver 580.126, the same method, the same fingerprints):
| Row (5090, stock) | MH/s | W | Energy per hash | Against w4 | Label |
|---|---|---|---|---|---|
| w4, full occupancy (four rows) | 142.5 to 142.7 | 311 to 319 | 2.21 microjoules | the control (the record's 2.26 at 311 W) | measured |
| w64 exported, full occupancy | 75.3 | 344 | 4.57 | -47 percent of rate (the 5 October row reproduced) | measured |
| w64 unhinted, best sparse sp170-w1 | 101.5 | 343 | 3.38 | -29 | measured |
| w64-l2 hinted, full occupancy | 110.4 | 392 | 3.55 | -22.5 | measured |
| **w64-l2 hinted, sp170-w4** | **145.7** | **401** | **2.75** | **+2.3 percent of rate, PASS; +25 percent of energy** | measured |
| w64-l2 hinted, sp170-w8 | 141.9 | 408 | 2.88 | -0.4; +30 percent | measured |
The 5090's second sector at stock: +88 W at 18.6 G reads a second = 4.7 nJ per sector (the 4090 8.6) against the DRAM's modelled 1.15, k about 0.24 on the sector. At zero shadow against the GDDR7 chip 2.21 / 0.466 = 4.7x becomes 2.75 / 0.62 = 4.4x (-7 percent); with the class v4 shadow at k 0.5, 3.3x unchanged (the shadow dilutes it); against the SRAM die 2.75 / 0.126 = 22x. Lane 3's reversal line was a sector under about 2 nJ; 4.7 is not, so the verdict holds on the 5090 too: W = 16 out of class v6 at stock, 0.3x at zero shadow and 0 with the shadow for +88 W on every 5090. The nuance for the row: on Blackwell the hinted form needs the sparse shape (full occupancy -22 percent, sp170-w4 +2), where Ada held it at full occupancy; so "the hint holds the 4-byte rate" is true on Ada and Blackwell with the right shape, false on Hopper and Ampere. The knee row stays owed to PC 1 as a formality (the lever lives only under about a 20 percent energy rise there, and stock reads 25). All five pods destroyed; raw logs under `~/igneum-fleet/fl5-results/`; spend about USD 9 of the 150.
## 8. Unverified and owed