Skip to content

FreeForGem NEB 1.0.10

Latest

Choose a tag to compare

@pepsykolya pepsykolya released this 21 Nov 02:17
· 1 commit to main since this release

FreeForGem NEB 1.0.10

For the NEB 1.0.1 main network. Recommended for everyone; faster on RTX 20 / GTX 16 / CMP Turing cards and on RTX 30. Mining, pools, dataset (NEB 1.0.1, BLAKE3 790b57fb…9e72), the automatic path choice, auto-lock and the dev fee (1%, on a connection of its own) are as in 1.0.9. GPU engine 0.6.1. auto-uv now tunes while the cards mine.

Faster nonce screen on the digest-table path. Engine benchmark on the same models: RTX 3070 +28.5–28.9 % rate (+8.6–11.5 % per watt), CMP 50HX (Turing) +15–21 % rate (+10–14 % per watt). Bit-exact.

  • Faster screen. On the digest-table path the engine first screens the nonces for the attempts with at most two live rows and then looks those up in the model's table. 0.6.1 does that screen faster. It is on by default on every card that takes the table path (CC 7.5 and newer: RTX 20 / GTX 16 / CMP Turing, RTX 30 / 40 / 50). GTX 10 and Volta keep the full path as before.
  • More power. The faster screen keeps the GPU busier, so the card draws more power: about +17 % on the RTX 3070 in the benchmark. The gain per watt is therefore a third to a half of the gain in rate.
  • auto-uv tunes while mining. A card under --auto-uv keeps mining on the path the engine chose, sends shares and counts its MA/s while its steps are measured (in 1.0.9 a card under tuning stood at 0 MA/s for its whole tuning). All GPUs are tuned at once. Each step's power is compared with the best offset so far on the same model and path, and a step ends as soon as the difference is clear. The rules are unchanged: climb while every step lowers the power by more than 0.5 W, resume 2 steps below a failure.
    • Measured on an RTX 3070: a fresh tuning took 120 s, including one fault and an automatic restart of the miner (about 17 s per step). A start from a saved profile needs one 17-s check. The card ended at +240 with 124–126 W, against 129–130 W at a fixed +210, at the same rate (about −3 %).
    • A hard Xid at the edge costs one self-restart of the miner of about 10 s. A whole rig tunes in roughly 2–5 min, all cards in parallel (estimate).
    • Tuning lines in the log are yellow, failures and roll-backs orange. FFGN_UV_CHILD=1 brings back the old separate check process.
  • When all GPUs stall at once after a CUDA fault on one GPU, the log now says so (instead of "likely a PCIe bus / power event").
  • Results, models, tickets and the dataset are unchanged bit for bit. The CPU reference still re-checks every share.

On the table path the evaluated rate (MA/s) = screened nonces × the model's share of attempts with at most two live rows. The screened rate (G/s) and the screened nonces per watt therefore compare the versions even when the chain's model changed between runs.

Engine benchmark, 1.0.9 engine (0.6.0) against 1.0.10 engine (0.6.1), same recorded live models, one process, same digest table (21.11.2026):

Card, core lock Model (share with ≤ 2 live rows) 1.0.9 engine 1.0.10 engine Rate MA/W
RTX 3070, 1680 MHz / +210 / 200 W Kepler (1.12 %) 4.97 G/s, 55.8 MA/s, 110.3 W, 0.506 MA/W 6.41 G/s, 71.9 MA/s, 127.5 W, 0.564 MA/W +28.9 % +11.5 %
RTX 3070, 1680 MHz / +210 / 200 W Kepler (1.56 %) 4.93 G/s, 77.1 MA/s, 112.0 W, 0.688 MA/W 6.34 G/s, 99.1 MA/s, 130.0 W, 0.762 MA/W +28.6 % +10.8 %
RTX 3070, 1680 MHz / +210 / 200 W TESS (3.92 %) 4.72 G/s, 185.3 MA/s, 113.0 W, 1.640 MA/W 6.07 G/s, 238.2 MA/s, 133.8 W, 1.781 MA/W +28.5 % +8.6 %
CMP 50HX (Turing), 1300 MHz Kepler (1.12 %) 4.64 G/s, 52.06 MA/s, 111.8 W, 0.466 MA/W 5.60 G/s, 62.85 MA/s, 118.0 W, 0.533 MA/W +20.7 % +14 %
CMP 50HX (Turing), 1300 MHz TESS (3.92 %) 4.38 G/s, 171.66 MA/s, 112.8 W, 1.522 MA/W 5.03 G/s, 197.48 MA/s, 117.9 W, 1.676 MA/W +15.0 % +10.1 %

Live mining on one pool, screened nonces per second and per watt (the chain's models differ between the runs):

Where Slot 1.0.9 1.0.10 Rate Per watt
RTX 3070, 1680 MHz / +210 / 200 W Kepler 4.95 G/s at 111.4 W 6.18 G/s at 132.8 W +25 % +4.8 %
RTX 3070, 1680 MHz / +210 / 200 W TESS 4.55 G/s at 117.5 W 5.81 G/s at 139.6 W +27 % +7.6 %
5× CMP 50HX (Turing), 1300 MHz, rig total Kepler 23.43 G/s at 567 W 26.51 G/s at 605 W +13.1 % +6.1 %
5× CMP 50HX (Turing), 1300 MHz, rig total TESS 21.09 G/s at 596 W 21.73 G/s at 604 W +3.0 % +1.6 %

Live runs of 1.0.10:

  • RTX 3070, first 16 min:
    • Kepler (model share 1.93 %): 186.6 MA/s at 132.8 W (1.405 MA/W).
    • TESS (8.62 %): 428.9 MA/s at 139.6 W (3.07 MA/W).
    • Shares: 63 found, 62 accepted, 1 stale ("job not found" right after a job change), 0 bad.
  • Five CMP 50HX, 20 min:
    • Kepler (2.77 %): 739.4 MA/s at 605 W for the rig (1.222 MA/W), 146.2–149.5 MA/s per card.
    • TESS (9.13 %): 1981.5 MA/s at 604 W (3.276 MA/W), 389.5–404.1 MA/s per card.
    • Shares: 117 found, 0 bad. At the pool: 84 accepted, 1 stale.

On TESS at a high model share, much of the GPU time goes to the attempts themselves, so the faster screen helps less there. We did not measure a desktop RTX 40 / 50 with 0.6.1. The RTX 4070 Laptop GPU's speed follows its thermal limit. No figure is claimed for either.

Note: all GPUs of a rig run in one miner process, so a fault on one GPU (usually an overclock or undervolt too far for that card) restarts the miner on every GPU; it recovers by itself in about a minute. The log names the GPU with the driver fault: lower that card's offset or use --auto-uv.

These are measurements, not a recommendation to change clocks. The release does not change your OC or auto-lock settings.

Verification: zero mismatches on Windows RTX 4070 Laptop, Linux RTX 3070 and Linux CMP 50HX (Turing). Tests run:

  • consensus vectors and s1, self-test 90/90, robust/random/scan on trained, live and synthetic models;
  • the new screen against the old one on the same nonces (all table hits identical, CPU checks);
  • automatic path against the full path (subset, identical tickets);
  • incremental table rebuilds, several engine contexts in one process;
  • cancel latency on large table calls (4–5 ms).

The packaged Windows miner passes --self-test 90/90 (RTX 4070 Laptop). The engine in the Linux packages has the same GPU code as the 1.0.10 engine that mined live on RTX 3070 with 0 bad results. The new auto-uv was tested live on an RTX 3070. Host tests: 144/144.

HiveOS: custom miner ffgn-hive, algorithm nebhash, wallet template %WAL%.%WORKER_NAME%, Installation URL:
https://github.com/pepsykolya/FreeForGemNeb/releases/download/v1.0.10/ffgn-hive-1.0.10.tar.gz
(flight sheet: HIVEOS.md). Guide: MANUAL.md.

The miner makes no promise about the value of the coin or any income.

SHA-256

9e50c811c6fc3a846943efd4ae9111dca88d5532f1272893bf6e5b468d17b736  ffgn-1.0.10-linux-x86_64.tar.gz
ef1b695af8af62ce817f211bbba5957e0bf7b030a4c43b13c0512acd72f6eda6  ffgn-hive-1.0.10.tar.gz
5d7ac3c03789cd8e5760dec84ad8e0e9b43c869a469cdba832b8392422c79217  ffgn-1.0.10-windows-x64.zip

По-русски

Для основной сети NEB 1.0.1. Рекомендуется всем; быстрее на картах RTX 20 / GTX 16 / CMP на Turing и на RTX 30. Майнинг, пулы, датасет NEB 1.0.1, автоматический выбор пути, auto-lock и комиссия (1%, отдельным соединением) — как в 1.0.9. GPU-движок 0.6.1. auto-uv теперь подбирает, пока карты майнят.

Быстрее проверка нонсов на пути таблицы дайджестов. Бенчмарк движка на тех же моделях: RTX 3070 — скорость +28,5…+28,9 % (на ватт +8,6…+11,5 %), CMP 50HX (Turing) — скорость +15…+21 % (на ватт +10…+14 %). Результаты бит в бит.

  • Быстрая проверка. На пути таблицы движок сначала проверяет нонсы, отбирая попытки, где живых строк не больше двух, и затем ищет их в таблице модели. В 0.6.1 эта проверка быстрее. Она включена по умолчанию на всех картах, которые идут путём таблицы (CC 7.5 и новее: RTX 20 / GTX 16 / CMP на Turing, RTX 30 / 40 / 50). GTX 10 и Volta, как и раньше, остаются на полном пути.
  • Больше мощность. Быстрая проверка сильнее загружает GPU, поэтому карта потребляет больше: на RTX 3070 в бенчмарке примерно +17 %. Выигрыш на ватт — от трети до половины выигрыша в скорости.
  • auto-uv подбирает на ходу. Карта под --auto-uv продолжает майнить выбранным движком путём, отправляет шары и считает свои MA/s, пока замеряются её шаги (в 1.0.9 карта во время подбора всё время стояла на 0 MA/s). Все карты подбираются одновременно. Мощность каждого шага сравнивается с лучшим смещением на той же модели и том же пути, и шаг заканчивается, как только разница ясна. Правила прежние: подъём, пока каждый шаг снижает мощность больше чем на 0,5 Вт, после сбоя — продолжение на 2 шага ниже.
    • Замер на RTX 3070: подбор с нуля занял 120 с вместе с одним сбоем и автоматическим перезапуском майнера (около 17 с на шаг). Старт с сохранённого профиля — одна проверка 17 с. Итог +240 при 124–126 Вт против 129–130 Вт на фиксированных +210 при той же скорости (около −3 %).
    • Жёсткий Xid на краю стоит одного самоперезапуска майнера примерно на 10 с. Весь риг подбирается примерно за 2–5 мин, все карты параллельно (оценка).
    • Строки подбора в логе жёлтые, сбои и откаты — оранжевые. FFGN_UV_CHILD=1 возвращает прежний отдельный проверочный процесс.
  • Если все карты встали разом после сбоя CUDA на одной карте, лог теперь так и пишет (вместо «likely a PCIe bus / power event»).
  • Результаты, модели, тикеты и датасет не изменились бит в бит. Эталон на CPU по-прежнему перепроверяет каждую шару.

На пути таблицы вычисленная скорость (MA/s) = проверенные нонсы × доля попыток модели, где живых строк не больше двух. Поэтому скорость проверки (G/s) и проверенные нонсы на ватт сравнивают версии, даже если модель цепи между прогонами сменилась.

Бенчмарк движка (те же записанные живые модели, один процесс, та же таблица; таблица выше):

  • RTX 3070, 1680 МГц / +210 / 200 Вт:
    • Kepler (1,12 %): 4,97 → 6,41 G/s, 55,8 → 71,9 MA/s, 0,506 → 0,564 MA/W;
    • Kepler (1,56 %): 77,1 → 99,1 MA/s, 0,688 → 0,762 MA/W;
    • TESS (3,92 %): 185,3 → 238,2 MA/s, 1,640 → 1,781 MA/W.
  • CMP 50HX, 1300 МГц:
    • Kepler: 52,06 → 62,85 MA/s (+20,7 %), 0,466 → 0,533 MA/W;
    • TESS: 171,66 → 197,48 MA/s (+15,0 %), 1,522 → 1,676 MA/W.

Живой майнинг, проверенные нонсы (1.0.9 → 1.0.10):

  • RTX 3070: Kepler 4,95 → 6,18 G/s (+25 %, на ватт +4,8 %), TESS 4,55 → 5,81 G/s (+27 %, на ватт +7,6 %).
  • Пять CMP 50HX при 1300 МГц: Kepler 23,43 → 26,51 G/s (+13,1 %, на ватт +6,1 %), TESS 21,09 → 21,73 G/s (+3,0 %, на ватт +1,6 %).

Живые прогоны 1.0.10:

  • RTX 3070, первые 16 мин: Kepler 186,6 MA/s при 132,8 Вт, TESS 428,9 MA/s при 139,6 Вт. Найдено 63 шары, принято 62, 1 устаревшая, неверных 0.
  • Пять CMP 50HX, 20 мин: Kepler 739,4 MA/s на риг при 605 Вт, TESS 1981,5 MA/s при 604 Вт. Найдено 117 шар, неверных 0.

На TESS при большой доле модели много времени GPU уходит на сами попытки, поэтому там быстрая проверка помогает меньше. Настольные RTX 40 / 50 с движком 0.6.1 мы не мерили. Скорость RTX 4070 Laptop GPU следует её тепловому лимиту. Цифр для них не заявляем.

Примечание: все карты рига работают в одном процессе майнера, поэтому сбой одной карты (обычно разгон или андервольт, слишком сильный для этой карты) перезапускает майнер на всех картах; он сам восстанавливается примерно за минуту. В логе названа карта со сбоем драйвера: уменьшите её смещение или используйте --auto-uv.

Это результаты замеров, а не рекомендация менять частоты: релиз не меняет ваши OC и auto-lock.

Проверки: 0 расхождений на Windows RTX 4070 Laptop, Linux RTX 3070 и Linux CMP 50HX (Turing). Что проверено:

  • обычные и s1-векторы, self-test 90/90, robust/random/scan на обученных, живых и синтетических моделях;
  • новая проверка против старой на тех же нонсах (все хиты таблицы совпадают, проверки на CPU);
  • автоматический путь против полного (подмножество, одинаковые тикеты);
  • инкрементальные пересборки таблицы, несколько контекстов движка в одном процессе;
  • задержка отмены на больших вызовах таблицы (4–5 мс).

Упакованный майнер для Windows проходит --self-test 90/90 (RTX 4070 Laptop). Движок в Linux-пакетах — тот же GPU-код, что у движка 1.0.10, майнившего вживую на RTX 3070 без неверных результатов. Новый auto-uv проверен вживую на RTX 3070. Тесты хоста: 144/144.

HiveOS: кастомный майнер ffgn-hive, алгоритм nebhash, Installation URL — см. выше. Подробности: README.ru.md, MANUAL.ru.md.