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

Counter ASIC 3.0 status: the RTX 4070 ladder (holds to 200,000 ops at its tune point; binds at 330,700 under the 160 W cap)

Browse source
This commit is contained in:
igneum-labs 2026-10-06 17:49:44 +00:00
parent 1fbdf1008e
commit 6a0d4eb59f
1 changed files with 13 additions and 0 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

13
docs/plans/counter-asic-3-status.md
View file

@ -88,6 +88,19 @@ Chip side (approximate: the f = 1 stored-dataset chip of item 1 plus an ALU core
Consequences per tier at N = 100,000: an Apple miner loses 1.5 percent of rate and pays 16 W more (0.56x per watt, per pound unchanged); a 5090 loses 0.2 percent and goes from 350 to 431 W (0.81x per watt), so a 5090 rig pays about 23 percent more electricity for the same blocks; a pool user sees nothing; a small NVIDIA card (4060 class) binds near 100,000 by its ALU budget (model, owed); AMD holds by its budget (owed); every verifier tier is untouched (+0.17 ms per warp).
### Item 8, the RTX 4070 rows (PC 1 job run-ca3-pc1-4070-shadow-20261006, 17:36:05 to 17:48:10Z, exit 0; the 4070 alone through api/cards, the installed CUDA worker's --bench, nvidia-smi at 1 Hz with 12 of 12 idle samples carrying power; the card at its Ember tune point: a 1,860 MHz core lock and the 160 W cap, memory 10,251 MHz; the 5090 and 9070 XT mining beside it; every pack's fingerprint equal to the Mac's)
| N ops per hash | Pack | 4070 MH/s (delta) | Watts (min to max) | Microjoules per hash | SM MHz | Reading |
|---|---|---|---|---|---|---|
| 930 (class v3) | mx8-devnet-epoch0, twice | 30.95 | 79.3 to 79.8 | 2.56 to 2.58 | 1,860 | the control: 2.5x the M5 Max's energy per hash, a third more than the 5090's at the hash |
| 49,700 | sh256x13 | 31.08 (+0.4%) | 93.5 | 3.01 | 1,860 | |
| 49,700, 64-instruction block | sh64x52 | 32.09 (+3.7%) | 92.6 | 2.89 | 1,860 | the 64-block gain seen on the Mac and the 5090 |
| 102,100 (the candidate) | sh256x27 | 31.08 (+0.4%) | 109.0 | 3.51 | 1,860 | latency-bound; +30 W for no rate |
| 199,600 | sh256x53 | 31.07 (+0.4%) | 138.2 | 4.45 | 1,860 | still latency-bound |
| 330,700 | sh256x88 | 27.14 (-12.3%) | 159.9 (the cap) | 5.89 | 1,846 | the 160 W cap binds; compute-bound under it |
Reading: the model's "a small NVIDIA card binds near 100,000" is replaced by the measurement: the 4070 holds its rate to about 200,000 ops per hash at its tune point and binds only at 330,700 when its power cap does. Consequence per tier: a 12 GB NVIDIA owner on a 4070 loses no rate at the candidate and pays 30 W more (79 to 109, 0.73x per watt); the N no card we own loses 5 percent at stays set by the M5 Max (about 130,000), not by the small card. One line to check: the restore printed "no entry was enabled before this job" and waited for no worker, while the card had been mining at 28.8 MH/s before it; the card's state after the job is read from the intake below.
### Item 8, the RX 9070 XT rows (PC 1 job run-ca3-pc1-amd-g1-shadow-20261006, 15:46:27 to 15:56:54Z, exit 0; the 9070 XT alone through api/cards in every key form (amd:3:gfx1201 is the installed key today), confirmed by the process list, restored in finally and mining again 15 s later; the 5090 and 4070 mining beside it, the 5090 probably clock-locked at about 2,781 MHz from the aborted Ember step; AMD OpenCL 3683.0, device-event time, 2^24 per dispatch)
| N ops per hash | Pack | 9070 XT MH/s (window s) | OpenCL build ms / dataset ms | Watts | Fingerprint = the Mac's |