Bench table: the model sweep's last rows (RTX 4070 Ti 31.24 MH/s at 95.4 W; L40S 56.41 at 220.7 W in the datacentre table; RTX 4080 Super 42.60 at 134.9 W); the GTX 1080 Ti row's flag cleared by its second host (five of five self-tests pass, the failure was the first host's VRAM); the page carries the sweep's two findings as one line (class v4 cap-proof; class v5 level on uncapped cards and 17 to 18 percent under on two power-capped hosts, so a v5 rate needs the power state beside it)
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
This commit is contained in:
igneum-labs
parent
09ce623eab
commit
443ac757fe
3 changed files with 77 additions and 7 deletions
|
|
@ -477,6 +477,8 @@ for (const [file, active] of PAGES) {
|
|||
const bestStock = ranked(cur.filter(r => tunedState(r)[0] === 'stock'))[0];
|
||||
const bestTuned = ranked(cur.filter(r => tunedState(r)[0] === 'tuned'))[0];
|
||||
const name = (r) => esc(r.card.replace(/ \(.*$/, ''));
|
||||
// the rented model sweep of 8 October 2026 (23 rows on 22 models, every class v5 fingerprint matched): its two page-level findings
|
||||
const sweepLine = `<p class="lede sweep">The rented model sweep of 8 October 2026 (23 rows, 22 models, every row untuned) found class v4's rate cap-proof: a card on a host power cap gives the same MH/s as an uncapped one. Class v5 held class v4's rate on every uncapped card and read 17 to 18 percent under it on the two power-capped hosts (an A4000 at 114 W, a GTX 1080 Ti at 130 W), so a class v5 rate needs the card's power state beside it; the table's rates are class v4.</p>`;
|
||||
const bestLine = best ? `<p class="lede"><strong>Best desktop card:</strong> ${name(best)}, ${esc(fmt(best.mh_s))} MH/s (measured, ${esc(best.date)}).`
|
||||
+ (bestStock ? ` <strong>Best stock MH per wall watt:</strong> ${name(bestStock)}, ${esc(fmt3(bestStock.mh_per_w))} at ${esc(fmt(bestStock.watts))} W, untuned (${esc(bestStock.by)}, ${esc(bestStock.date)}).` : '')
|
||||
+ (bestTuned ? ` <strong>Best tuned:</strong> ${name(bestTuned)}, ${esc(fmt3(bestTuned.mh_per_w))} at ${esc(fmt(bestTuned.watts))} W (measured, ${esc(bestTuned.date)}).` : '') + '</p>' : '';
|
||||
|
|
@ -553,6 +555,7 @@ table.bench2 tr.detail .d{display:block;margin:0 0 4px}table.bench2 tr.detail b{
|
|||
sortStyle,
|
||||
'<h2 id="table">The table</h2>',
|
||||
bestLine,
|
||||
sweepLine,
|
||||
'<p>One row per card on the current class: the class v4 program (the latency-shadow block over the class v3 hash), or a class v3 row re-measured with its class v4 cost on 6 October 2026 or later. MH per wall watt uses board or wall power; a row whose watts are the chip\'s (Apple silicon: GPU plus DRAM from IOReport) says so and is not ranked on that column. Integrated GPUs are not listed.</p>',
|
||||
'<h3 class="grp" id="buy">Cards you can buy</h3>',
|
||||
table,
|
||||
|
|
|
|||
|
|
@ -1145,8 +1145,8 @@
|
|||
"watts": 124.3,
|
||||
"mh_per_w": 0.131,
|
||||
"driver_os": "NVIDIA driver 580.173.02, Ubuntu 24.04 (CUDA 12.8.1 image)",
|
||||
"source": "model sweep 2026-10-08, row 17 (a rented card, result cb2-gtx-1080-ti-11-gb-0377-result.json; 09:49Z)",
|
||||
"note": "FLAG: one of three class v4 builds on this host failed its cache self-test (the 256 MiB cache's FNV read wrong on that build while the other two and the class v5 kit build were clean: the host's VRAM or a Pascal-only build race); a second host decides, and until then this row carries the flag. The rate is the mean of the two clean runs (16.261, 16.260). Rented single-card host, 0.10 h; power.draw mean 124.3 W (max 125.5, limit 280 W; no power average on Pascal), SM 1,936 MHz mean, memory 5,005 MHz, 59 C; self-test PASS on the clean builds (96 of 96 vector lanes); the class v5 kit holds the rate (16.27 MH/s, fingerprint matched); read ceiling 16.3 MH/s, the hash at 1.00 of it. The first Pascal row: a 2017 card that runs the class at its own ceiling",
|
||||
"source": "model sweep 2026-10-08, rows 17 and 22 (two rented hosts, results cb2-gtx-1080-ti-11-gb-0377-result.json and cb2-gtx-1080-ti-11-gb-1452-result.json; 09:49Z and 10:02Z)",
|
||||
"note": "the mean of the first host's two clean runs (16.26 MH/s at 124.3 W, the card uncapped at a 280 W limit); a second host read 16.29 MH/s at 129.4 W under a 130 W host cap with five of five self-tests passing, so the first host's one-in-three cache self-test failure was that host's VRAM and not Pascal, and the flag is cleared. Power.draw on the first host max 125.5 W (no power average on Pascal), SM 1,936 MHz mean, memory 5,005 MHz, 59 C; the class v5 kit held the rate on the uncapped host (16.27 MH/s) and read 13.27 MH/s on the 130 W-capped one (fingerprint matched on both); read ceiling 16.3 MH/s, the hash at 1.00 of it. The first Pascal row: a 2017 card that runs the class at its own ceiling",
|
||||
"hive": {
|
||||
"core_mhz": null,
|
||||
"mem_mhz": null,
|
||||
|
|
@ -1197,6 +1197,72 @@
|
|||
"pl_w": null,
|
||||
"label": "stock (rented card, no clock control on the host; no measured tune point)"
|
||||
}
|
||||
},
|
||||
{
|
||||
"generator": "v2",
|
||||
"miner": "bench pack, igneum-worker-cuda, class v4 program (3 x 300 dispatches of 2^22)",
|
||||
"date": "2026-10-08",
|
||||
"by": "measured by the fleet",
|
||||
"v4_cost": "not measured (class v4 program only)",
|
||||
"tuned": "stock, bench only (rented, no clock control on the host)",
|
||||
"card": "NVIDIA RTX 4070 Ti (12 GB)",
|
||||
"group": "buy",
|
||||
"mh_s": 31.24,
|
||||
"watts": 95.4,
|
||||
"mh_per_w": 0.327,
|
||||
"driver_os": "NVIDIA driver 580.126.09, Ubuntu 24.04 (CUDA 12.8.1 image)",
|
||||
"source": "model sweep 2026-10-08, row 20 (a rented card, result cb2-rtx-4070-ti-12-gb-0695-result.json; 09:57Z)",
|
||||
"note": "rented single-card host, 0.18 h; power.draw mean 95.4 W (max 96.7, limit 285 W), SM 2,825 MHz mean, memory 10,251 MHz, 49 C, no throttle reason; self-test PASS (96 of 96 vector lanes); the class v5 kit holds the rate (31.28 MH/s, fingerprint matched); read ceiling 31.7 MH/s, the hash at 0.99 of it. Repeats the 7 October rate to 0.07 percent on another provider at 12 W less; the 192-bit Ada pair (4070 Ti, 4070 Super) land on one number, 31.2 MH/s under 100 W",
|
||||
"hive": {
|
||||
"core_mhz": null,
|
||||
"mem_mhz": null,
|
||||
"pl_w": null,
|
||||
"label": "stock (rented card, no clock control on the host; no measured tune point)"
|
||||
}
|
||||
},
|
||||
{
|
||||
"generator": "v2",
|
||||
"miner": "bench pack, igneum-worker-cuda, class v4 program (3 x 300 dispatches of 2^22)",
|
||||
"date": "2026-10-08",
|
||||
"by": "measured by the fleet",
|
||||
"v4_cost": "not measured (class v4 program only)",
|
||||
"tuned": "stock, bench only (rented, no clock control on the host)",
|
||||
"card": "NVIDIA L40S (48 GB)",
|
||||
"group": "datacentre",
|
||||
"mh_s": 56.41,
|
||||
"watts": 220.7,
|
||||
"mh_per_w": 0.256,
|
||||
"driver_os": "NVIDIA driver 570.133.20, Ubuntu 24.04 (CUDA 12.8.1 image)",
|
||||
"source": "model sweep 2026-10-08, row 21 (a rented card, result cb2-l40s-48-gb-0890-result.json; 09:57Z)",
|
||||
"note": "rented single-card host, 0.17 h; power.draw mean 220.7 W (max 227.2, limit 350 W), SM 2,520 MHz, memory 9,001 MHz, 45 C, no throttle reason; self-test PASS (96 of 96 vector lanes); the class v5 kit holds the rate (56.51 MH/s, fingerprint matched); read ceiling 57.8 MH/s, the hash at 0.98 of it. Repeats the 7 October row (56.36 then, 56.41 now) at 20 W less on this host; a 384-bit GDDR6 card that lands where the 3090 Ti and the A6000 do at a datacentre price. Driver 570.133, the oldest of the sweep, ran the CUDA 12.8 worker clean: 570 is a safe floor",
|
||||
"hive": {
|
||||
"core_mhz": null,
|
||||
"mem_mhz": null,
|
||||
"pl_w": null,
|
||||
"label": "stock (rented card, no clock control on the host; no measured tune point)"
|
||||
}
|
||||
},
|
||||
{
|
||||
"generator": "v2",
|
||||
"miner": "bench pack, igneum-worker-cuda, class v4 program (3 x 300 dispatches of 2^22)",
|
||||
"date": "2026-10-08",
|
||||
"by": "measured by the fleet",
|
||||
"v4_cost": "not measured (class v4 program only)",
|
||||
"tuned": "stock, bench only (rented, no clock control on the host)",
|
||||
"card": "NVIDIA RTX 4080 Super (16 GB)",
|
||||
"group": "buy",
|
||||
"mh_s": 42.6,
|
||||
"watts": 134.9,
|
||||
"mh_per_w": 0.316,
|
||||
"driver_os": "NVIDIA driver 580.126.09, Ubuntu 24.04 (CUDA 12.8.1 image)",
|
||||
"source": "model sweep 2026-10-08, row 23 (a rented card, result cb2-rtx-4080-super-16-g-2453-result.json; 10:03Z)",
|
||||
"note": "rented single-card host, 0.12 h; power.draw mean 134.9 W (max 138.4, limit 320 W), SM 2,740 MHz mean, memory 11,254 MHz, 51 C, no throttle reason; self-test PASS (96 of 96 vector lanes); the class v5 kit holds the rate (42.68 MH/s, fingerprint matched); read ceiling 43.1 MH/s, the hash at 0.99 of it. The 23 Gbps GDDR6X gives it 4.6 percent over the 4080 for 6 W more at the same MH per watt: the Ada 256-bit tier is 41 to 43 MH/s whichever card",
|
||||
"hive": {
|
||||
"core_mhz": null,
|
||||
"mem_mhz": null,
|
||||
"pl_w": null,
|
||||
"label": "stock (rented card, no clock control on the host; no measured tune point)"
|
||||
}
|
||||
}
|
||||
]
|
||||
}
|
||||
|
|
|
|||
File diff suppressed because one or more lines are too long
Loading…
Reference in a new issue