14 KiB
The reference GPU population and the cost per accepted unit of work
8 October 2026, 18:0x UK, branch class-v6-floor-sram, floor lane 3 for the research lane (the pin appended to
docs/plans/igneum-2.0.md by the coordinator at 18:16 UK: the served reference population, one row per class, two tests
per class, every cell labelled). Beside docs/analysis/class-v6/coexistence-model.md (its section 2 is most of this)
and the workbook. Every energy cell carries its label (measured or modelled or estimated, with where); every cost cell
is arithmetic on labelled inputs and is modelled; prices are approximate street and list figures of October 2026.
The script is scratchpad/refpop.py, run on build-4 at 18:02 UK. Apple is reported and not headlined. Nothing here is
a boundary; the founder is not named.
1. The measure
Cost per accepted MH/s-hour, in micro-USD (USD x 10^-6): annualised hardware plus power plus hosting, failures and fees, divided by the year's accepted work. Accepted work is 97 percent of raw (rejects 0.5 percent, downtime 2.0, epoch preparation and propagation 0.5; approximate from the devnet's share rates); the pool fee 1 percent. Two tests per class:
- The existing owner: hardware sunk; pays power at the card's floor joules (the knee with the knobs, lane 4's
class v5 table and the tiers file
app/igneum-app/tiers/class-v5-tiers.json), a wear allowance of 5 percent of the used price a year, and fees; the alternative use (the card's rental yield on a public market, approximate) is reported in its own column and not deducted, so the owner's number is the floor of what the card costs to run. - The new entrant: buys at MSRP (the list price) or at street (October 2026, approximate), operates two years, resells at the table's fraction of MSRP, with 3 percent a year of failures; power as the owner.
Three electricity prices (USD 0.06, 0.12, 0.25 per kWh); hosting 0 at home. The energy columns: the floor is the knee with the card's knobs (the lock on Blackwell and Ada, the cap on Ampere, the ADLX offsets on AMD; nothing on Apple or Intel), the stock row unlocked at 100 percent; "measured" names the instrument's row (PC 1, PC 2, a rented host, lane 4's sweep), "modelled" a knee read from a measured stock row by the card's architecture, "estimated" a figure with no instrument behind it (the Arc B580's watts).
2. The plan's population (block 1)
| Class | Memory | Microjoules per hash, floor / stock | MH/s at the floor | Wh per MH/s-hour | MSRP / street / used (USD) | Resale after 2 y | Existing owner at 0.06 / 0.12 / 0.25 | New entrant at MSRP, 0.06 / 0.12 / 0.25 | New entrant at street, 0.06 / 0.12 / 0.25 | Hardware share of the street entrant at 0.12 | Alternative use (rental yield), per MH/s-hour | Energy label | Rate label |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| RTX 5090 | 32 GB | 2.33 / 3.48 | 134.8 | 2.33 | 1999 / 2600 / 2200 | 55 percent | 243 / 388 / 704 | 566 / 711 / 1027 | 846 / 992 / 1307 | 71 pct | 2967 | measured (PC 1, tiers file) | measured |
| RTX 5080 | 16 GB | 2.06 / 3.48 | 71.2 | 2.06 | 999 / 1100 / 900 | 50 percent | 204 / 332 / 611 | 570 / 699 / 978 | 660 / 788 / 1067 | 67 pct | 2809 | measured (rented, lane 4) | measured |
| RTX 4090 | 24 GB | 3.58 / 5.00 | 58.0 | 3.58 | 1599 / 1700 / 1300 | 45 percent | 357 / 580 / 1065 | 1178 / 1402 / 1887 | 1288 / 1512 / 1996 | 70 pct | 6034 | stock measured, knee modelled (lane 4) | measured |
| RTX 3090 (used) | 24 GB | 4.51 / 5.03 | 37.8 | 4.51 | 1499 / 900 / 650 | 30 percent | 384 / 666 / 1276 | 2030 / 2311 / 2922 | 1032 / 1314 / 1924 | 57 pct | 5291 | stock measured, cap modelled (lane 4) | measured |
| RX 9070 XT | 16 GB | 7.90 / 10.70 | 18.9 | 7.90 | 599 / 650 / 520 | 45 percent | 657 / 1151 / 2220 | 1591 / 2085 / 3154 | 1761 / 2255 / 3324 | 56 pct | 7937 | measured (PC 1, ADLX -500 MHz -30 pct) | measured |
| RX 7600 8 GB | 8 GB | 5.4 (modelled knee) / 7.2 | 15.8 | 5.40 | 269 / 270 / 200 | 40 percent | 413 / 750 / 1,481 | 981 / 1,318 / 2,049 | 985 / 1,322 / 2,053 | 49 pct | 3,167 | class v6: rate MEASURED 15.79 MH/s (PC 1, the AMD-and-Intel lane); stock 113 W measured, 7.2 microjoules; the knee 5.4 MODELLED by the 9070 XT's ADLX grid (24 percent of the draw for no rate), the metered knee the morning's batch 03; the earlier row (13.9 MH/s, 6.2 / 8.1) was the class v4 sub-version 3 read | measured rate and stock watts, modelled knee |
| Intel Arc B580 | 12 GB | 10.40 / 10.40 | 10.7 | 10.40 | 249 / 260 / 200 | 40 percent | 761 / 1411 / 2819 | 1529 / 2179 / 3587 | 1594 / 2243 / 3651 | 42 pct | 4673 | rate measured 10.4 to 11 MH/s (PC 2); watts unread, about 110 W by the board's class (estimated); no lever | measured rate, estimated watts |
| Apple M5 Max (reported, not headlined) | 36 GB | 1.40 / 1.40 | 27.1 | 1.40 | 3999 / 4000 / 3200 | 55 percent | 789 / 876 / 1066 | 4268 / 4356 / 4545 | 4271 / 4358 / 4548 | 96 pct | 0 (a workstation; an owner only) | measured (GPU and DRAM channels, 6 October) | measured |
Two more Apple rows (added 18:2x UK; the founder's own powermetrics reads, 8 October 18:1x UK, v2.0.0 class v5 on the 2.0 devnet, TEAM-REPORTED, GPU power only and not the wall, which is why they sit beside and not in place of the metered 1.40 row above; reported, not headlined; the Apple tier mines and does not prove; a Mac is an owner only, the entrant rows are the arithmetic for completeness):
| Class | Memory | Microjoules per hash (GPU power only) | MH/s | Wh per MH/s-hour | MSRP / street / used (USD) | Resale after 2 y | Existing owner at 0.06 / 0.12 / 0.25 | New entrant at MSRP, 0.06 / 0.12 / 0.25 | New entrant at street | Hardware share of the entrant at 0.12 | Energy label | Rate label |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| MacBook Pro M5 Max, 40 GPU cores, 64 GB | 64 GB unified | 0.73 | 22.67 | 0.73 | 3,999 / 4,000 / 3,200 | 55 percent | 884 / 930 / 1,028 | 5,043 / 5,089 / 5,188 | 5,046 / 5,092 / 5,191 | 98 percent | team-reported (powermetrics, GPU rail only, 16.65 W) | team-reported |
| Mac mini M6, 12 GPU cores, 16 GB | 16 GB unified | 0.73 | 11.0 | 0.73 | 999 / 1,000 / 800 | 55 percent | 478 / 523 / 622 | 2,619 / 2,664 / 2,763 | 2,624 / 2,670 / 2,769 | 97 percent | team-reported (powermetrics, GPU rail only, 8.03 W) | team-reported |
Reading: at the GPU rail alone both Macs are the most efficient honest joule in the table (0.73 against the 5070 Ti's 1.70), and still the dearest honest owners per accepted unit after the datacentre parts (the wear allowance on a USD 3,200 to 800 machine at 11 to 23 MH/s), because the Mac's cost is the machine, not the power; the M6 mini's owner (478 at 0.06) sits between the 3060 and the 9070 XT, the M5 Max laptop's (884) above every consumer card. The wall and the SoC's other rails are not in the 0.73; the metered M5 Max row (1.40 at the GPU and DRAM channels) is the comparable figure for the coexistence model.
3. The rest of the measured classes (block 2)
| Class | Memory | Microjoules per hash, floor / stock | MH/s at the floor | Wh per MH/s-hour | MSRP / street / used (USD) | Resale after 2 y | Existing owner at 0.06 / 0.12 / 0.25 | New entrant at MSRP, 0.06 / 0.12 / 0.25 | New entrant at street, 0.06 / 0.12 / 0.25 | Hardware share of the street entrant at 0.12 | Alternative use (rental yield), per MH/s-hour | Energy label | Rate label |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| RTX 5070 Ti | 16 GB | 1.70 / 2.84 | 77.0 | 1.70 | 749 / 800 / 650 | 50 percent | 156 / 263 / 493 | 412 / 519 / 749 | 454 / 560 / 790 | 62 pct | 1948 | stock measured (rented), knee modelled | measured |
| RTX 5070 | 12 GB | 1.75 / 2.99 | 41.0 | 1.75 | 549 / 560 / 450 | 50 percent | 175 / 284 / 521 | 531 / 640 / 877 | 548 / 657 / 894 | 67 pct | 2439 | stock measured, knee modelled | measured |
| RTX 5060 Ti 16 GB | 16 GB | 2.36 / 4.02 | 19.0 | 2.36 | 429 / 450 / 360 | 45 percent | 260 / 407 / 727 | 929 / 1077 / 1396 | 999 / 1146 / 1466 | 74 pct | 4211 | stock measured, knee modelled | measured |
| RTX 5060 | 8 GB | 2.21 / 3.76 | 17.0 | 2.21 | 299 / 310 / 250 | 45 percent | 225 / 364 / 663 | 747 / 885 / 1184 | 788 / 926 / 1225 | 70 pct | 3529 | stock measured, knee modelled | measured |
| RTX 4080 | 16 GB | 3.51 / 4.92 | 45.0 | 3.51 | 999 / 1000 / 700 | 40 percent | 312 / 531 / 1006 | 1058 / 1277 / 1752 | 1059 / 1279 / 1754 | 66 pct | 4444 | stock measured, knee modelled | measured |
| RTX 4070 | 12 GB | 3.58 / 5.82 | 31.1 | 3.58 | 549 / 550 / 400 | 40 percent | 300 / 524 / 1008 | 891 / 1114 / 1599 | 893 / 1116 / 1601 | 60 pct | 3215 | measured (PC, 1,860 MHz + 50 pct cap) | measured |
| RTX 4060 Ti 16 GB | 16 GB | 3.81 / 5.32 | 17.6 | 3.81 | 499 / 420 / 290 | 35 percent | 336 / 574 / 1090 | 1398 / 1636 / 2152 | 1116 / 1354 / 1870 | 65 pct | 4545 | stock measured, knee modelled | measured |
| RTX 3080 (used) | 10 GB | 4.20 / 4.54 | 40.8 | 4.20 | 699 / 450 / 330 | 25 percent | 310 / 573 / 1141 | 1071 / 1334 / 1902 | 687 / 950 / 1518 | 45 pct | 2941 | modelled (both rented hosts capped) | measured |
| RTX 3060 (used) | 12 GB | 5.77 / 6.40 | 23.8 | 5.77 | 329 / 260 / 190 | 25 percent | 408 / 768 / 1550 | 1013 / 1374 / 2155 | 831 / 1191 / 1972 | 39 pct | 2521 | stock measured, cap modelled | measured |
| H100 (hosted) | 80 GB | 2.00 / 2.58 | 90.0 | 2.00 | 25000 / 25000 / 18000 | 50 percent | 1313 / 1438 / 1709 | 8869 / 8994 / 9265 | 8869 / 8994 / 9265 | 97 pct | 27778 | stock measured (rented), lock modelled | measured |
| A100 (used) | 80 GB | 2.89 / 2.99 | 60.0 | 2.89 | 10000 / 10000 / 6000 | 35 percent | 775 / 955 / 1346 | 7001 / 7182 / 7573 | 7001 / 7182 / 7573 | 95 pct | 16667 | stock measured (rented) | measured |
4. The chip rows beside them, at the adversary lane's reconciled tickets (development sunk, farm hosting USD 0.02 per kWh on top)
| Chip (reconciled ticket) | Microjoules per hash | USD per MH/s | Life | All-in at 0.06 / 0.12 / 0.25 | Hardware / power at 0.06 | Label |
|---|---|---|---|---|---|---|
| GDDR7 machine (adversary lane, reconciled) | 1.55 | 4.84 | 1 y | 722 / 819 / 1029 | 586 / 128 | modelled: 1.5x per joule vs the 5090 lock, USD 4.84 per MH/s |
| GDDR7 machine (adversary lane, reconciled) | 1.55 | 4.84 | 3 y | 338 / 435 / 646 | 207 / 128 | modelled: 1.5x per joule vs the 5090 lock, USD 4.84 per MH/s |
| GDDR7 machine (adversary lane, reconciled) | 1.55 | 4.84 | 5 y | 262 / 359 / 569 | 131 / 128 | modelled: 1.5x per joule vs the 5090 lock, USD 4.84 per MH/s |
| Stored-half hybrid (reconciled) | 1.21 | 2.80 | 1 y | 443 / 519 / 682 | 339 / 100 | modelled: 1.93x, USD 2.80 |
| Stored-half hybrid (reconciled) | 1.21 | 2.80 | 3 y | 221 / 297 / 460 | 120 / 100 | modelled: 1.93x, USD 2.80 |
| Stored-half hybrid (reconciled) | 1.21 | 2.80 | 5 y | 177 / 253 / 416 | 76 / 100 | modelled: 1.93x, USD 2.80 |
| N2 SRAM die (reconciled, 1 to 1.5 kW machine) | 1.01 | 1.00 | 1 y | 207 / 270 / 407 | 121 / 84 | modelled: 2.3x, USD 1.0 (0.5 to 1.6) |
| N2 SRAM die (reconciled, 1 to 1.5 kW machine) | 1.01 | 1.00 | 3 y | 128 / 191 / 328 | 43 / 84 | modelled: 2.3x, USD 1.0 (0.5 to 1.6) |
| N2 SRAM die (reconciled, 1 to 1.5 kW machine) | 1.01 | 1.00 | 5 y | 112 / 175 / 312 | 27 / 84 | modelled: 2.3x, USD 1.0 (0.5 to 1.6) |
5. Reading the table
- On every consumer card the owner's cost is 60 to 70 percent electricity, so the grid price is the owner's whole variable; the entrant's cost is 60 to 75 percent hardware at street prices, so the card's price and its resale are the entrant's. The cheapest honest owner is a 5070 Ti at its knee (156 micro-USD at 0.06, 263 at 0.12); the cheapest entrant the same card at MSRP (412 at 0.06, 519 at 0.12).
- The 8 GB tier (the RX 7600 at 13.9 MH/s and 6.2 microjoules at its estimated knee, 8.1 measured at stock) costs an owner 2.6x the 5070 Ti per accepted unit and an entrant 1.4x at street; the Arc B580 (10.7 MH/s measured, watts estimated) 4x and 1.5x. These are the classes the dataset floor retires first, and their owners mine only below about 8 cents per kWh at the GPU equilibrium of the coexistence model.
- Datacentre parts and the Mac never enter as entrants (hardware 95 percent or more of the cost) and mine only as owners with nothing better to do, which their rental yields (USD 1 to 3 an hour, 10x to 30x their mining cost per MH/s-hour) say they always have.
- Against the chip rows at the reconciled tickets and a 3-year life, the GDDR7 machine (338 micro-USD at 0.06) sits above every Blackwell owner and between the Blackwell entrant at MSRP and at street; the hybrid (221) and the die (128) sit below every owner at 0.12. The coexistence model carries what that means; this table carries the cells.
6. Unverified and owed
- The RX 7600's knee is estimated by the 9070 XT's ADLX grid; its own 24-point grid runs on PC 1 and replaces the floor cell. Its dataset rows landed (18:2x UK, PC 1, rate only, fingerprints equal to the 5090's): 15.75 / 15.46 / 15.34 / 15.35 MH/s at 1 / 2 / 4 / 5.5 GiB, the 5.5 GiB dataset fitting at 6,732 MiB of 8,176; the table's 13.9 MH/s is the class v5 row at stock from the card-in job, within the day's spread (13.4 to 15.8 across jobs and classes); the 8 GB tier pays under 3 percent of rate for the 5.5 GiB step. The Arc B580's watts are unread (about 110 W by the board's class); the Intel lane's meter row replaces them. The Ada and Ampere knees are modelled (no rented host allows the lock).
- MSRP and street prices are October 2026 approximations; the 5090's street has been 2x MSRP this year; used and resale fractions are approximate.
- The rental yields are approximate public-market figures; the accepted-work factor, the wear allowance and the failure rate are assumptions.
- Nothing was run on the Mac; the script ran on build-4.