Bitcoin Forum
November 20, 2026, 02:39:12 AM *
News: Latest Bitcoin Core release: 32.1 [Torrent]
 
   Home   Help Search Login Register More  
Pages: « 1 [2]  All
  Print  
Author Topic: ForgeMiner (NVIDIA + AMD) — PRL · QUAN · NOID · AND 11 MORE COINS  (Read 766 times)
fenomenyaa
Jr. Member
*
Offline

Activity: 160
Merit: 2


View Profile
October 30, 2026, 05:57:24 PM
 #21

Hi,

Just wanted to update you that adding --kernel 2 completely fixed the issue.
I really appreciate your work on Forge Miner. The low CPU usage and efficient system resource management are fantastic—it feels like a breath of fresh air coming from other miners.

Looking forward to the upcoming updates and future releases. Thanks again for your hard work and great software!
0xHashRaptor (OP)
Jr. Member
*
Offline

Activity: 39
Merit: 5


View Profile
November 05, 2026, 06:31:46 PM
Last edit: November 18, 2026, 04:17:36 AM by 0xHashRaptor
 #22

Hi,

Just wanted to update you that adding --kernel 2 completely fixed the issue.
I really appreciate your work on Forge Miner. The low CPU usage and efficient system resource management are fantastic—it feels like a breath of fresh air coming from other miners.

Looking forward to the upcoming updates and future releases. Thanks again for your hard work and great software!

Thank you for the update and the kind words — great to hear that --kernel 2 sorted it out! In 1.9.1 that case is handled automatically: on RTX 30/40 with a driver older than 580, Pearl now falls back to --kernel 2 on its own

ForgeMiner v1.9.1 — fixes for what 1.9.0 users reported, Pearl faster on the RTX 50 series, NVIDIA core voltage on every RTX 20–50 card, --mclk with any clock the card lists, a fan curve per card, and a long round of fixes to pools, shutdown and the overclock flags.
  • Fixed: Pearl on the RTX 30 / 40 series with a driver older than 580 — the default kernel made the miner restart again and again. It now loads on those drivers, and falls back to --kernel 2 on its own where it cannot
  • Fixed: Pearl on the CMP 30HX mined at 0 H/s in 1.9.0; it now has a kernel of its own
  • Fixed: KawPow on the RX 470–590 — from epoch 613 on, the DAG build could hang the card
  • Fixed: --fan-curve stayed on its first point when the curve was written with spaces (40:30, 60:60)
  • Pearl on the RTX 50 series: up to +2% hashrate on the same overclock. Iron Fish on the RX 470–590: about +8%. Quantus on AMD: +0.3…1% on every series
  • NVIDIA core voltage on Linux: --vddc and --vddoff now work on every RTX 20 / 30 / 40 / 50 card. A card a request does not fit is skipped and the other cards are still set; the reason is shown in the dashboard
  • --mclk on the GTX 10 / 16, RTX 20 and CMP 30–50HX: any memory clock the card lists, not only 810. A clock the card does not list is refused with the list; if the driver accepts a memory clock but does not hold it, the miner says so instead of reporting success
  • A fan curve per card: --fan-curve 45:30,70:90/40:40,70:80/* — one curve per card in --gpu order, * leaves the card to the driver
  • --cclk above a card's highest clock now says what the driver will actually run; --oc-reset works per card on mixed rigs and with --gpu all; AMD fans go back to their own speed on exit
  • Pools: two failed connections move to the next pool in the list, and the first pool is taken back when it returns. A pool that hangs up on every full-size Pearl share gets one clear warning to mine on another pool
  • Every way out — Ctrl+C, miner stop, an error — hands the cards back and ends within seconds. Quantus: a card that hits a CUDA error restarts by itself while the other cards keep mining
  • NOID dev fee is now 2% (was 2.5%)
  • Windows rigs without an AMD card no longer load the AMD OpenCL runtime; on Linux, AMD cards hidden from the user running the miner get a clear "run as root" message; HiveOS: a --gpu in a trailing comment no longer applies, and backup pools with a scheme in the URL work

ForgeMiner v1.9.2 — Quantus faster on the RTX 30 / 40 / 50, KawPow faster on the GTX 10 and RTX 20–50, NVIDIA voltage fixed on many cards and now shown in the dashboard, RX 7000 cards in chipset slots no longer hang, and a card that fails while mining comes back by itself on every coin.
  • Quantus on the RTX 30 / 40 / 50 series: up to +1.6% on the same overclock
  • KawPow (Ravencoin, Quai, Neurai) on the GTX 10 and RTX 20 / 30 / 40 / 50 series: up to +5.5% on the RTX 30, up to +3.7% on the RTX 20, about +3% on the GTX 10, up to +1% on the RTX 40 / 50, on the same overclock
  • --vddoff / --vddc fixed on many RTX 20 / 30 / 40 / 50 cards: some cards refused the request ("Ambiguous voltage offset record") and mined at their own voltage. The miner now takes the voltage controls from the card itself
  • The current core voltage in the dashboard and the API (Linux): a Volt column next to Power, and voltage_mv for every card in the API. Green — lowered and confirmed on the card; yellow — requested but not applied. Lock the core clock and offset as usual (HiveOS or --cclk / --coff), then add --vddoff: the same hashrate for less power
  • Memory clock (Linux, NVIDIA driver 595 or newer): --mclk 405 now holds on the RTX 20 / 30 / 40 / 50, and --mclk 5000 / 810 / 405 on the CMP 90HX and CMP 50HX. On exit the memory goes back exactly as it was, a clock set in HiveOS included; after a crash the next start puts it back before mining
  • Pearl on the CMP 90HX at 405 MHz: about 27% less power for 1.6% less hashrate (12 cards). On the RTX 30 / 40 / 50 the Pearl difference is small: up to 1.6 W less per card and up to 0.9% less hashrate. Pearl on the RTX 20 and the CMP 50HX needs fast memory — at 810 MHz and below it loses most of its hashrate
  • A card that fails while mining comes back by itself on every coin (in 1.9.1 only Quantus did this): it is opened again after 10 s, 30 s, 1, 2, 5 and 10 minutes while the other cards keep mining. Conflux, Cryptix, Xelis and QubitCoin now restart a card whose launches keep failing instead of leaving it idle
  • Windows: forge.exe keeps the miner going — when a card cannot be brought back inside the running miner, the miner starts itself again, with a wait that doubles with each restart in a row, up to an hour. Ctrl+C, Ctrl+Break and closing the window still stop everything and hand the cards back. On HiveOS the miner exits and HiveOS starts it again, as before
  • RX 7000 in a slot without PCIe atomics (the chipset ports, where risers usually sit): Conflux, Xelis and KawPow no longer hang the card until a reboot
  • CMP unlock (--cmp-install) works on rigs with little memory — checked on a 4 GB rig with 12 CMP 90HX while it mined. On HiveOS, a rig that cannot reach the HiveOS server while it boots after the install no longer stays in maintenance mode, and --cmp-rollback restores the stock driver after a repeated --cmp-install

ForgeMiner v2.0.0 — dev fee down to 1% on Pearl, Quantus and NOID, Pearl and Quantus faster on NVIDIA and AMD, NOID about 39% faster on the RTX 20, temperature limits for the core and the memory separately, and two new coins — Zano and Compute Substrate.
  • Dev fee: Pearl, Quantus and NOID now take a 1% dev fee (was 2%), on NVIDIA and AMD
  • Pearl on the RX 9000 series (RDNA 4): +14.7% at the same power (RX 9060 XT)
  • Pearl on the RTX 30 / 40 / 50: up to +1.5% at the same power (RTX 5060 Ti +1.5%, RTX 4070 Ti +0.5%), plus fewer Invalid shares and less work lost at a block change
  • Quantus on the RTX 30: +0.9%, on the RTX 20 and CMP 40HX / 50HX: +0.7%, on AMD RX 6000 / 7000 / 9000 up to +0.2%
  • NOID on the RTX 20: about +39% (RTX 2080 Ti), on the RTX 40 +1.3%
  • New coin: Zano (ZANO) — --algorithm zano, NVIDIA and AMD, 0.7% dev fee. Ready-made launchers for LuckyPool (EU, Asia, Canada)
  • New coin: Compute Substrate (CSD) — --algorithm csd, NVIDIA and AMD, 1% dev fee. Ready-made launchers for LuckyPool (EU, RU)
  • RTX 50 series: the fastest Pearl kernel is now the default. --kernel 1 (default) is the fastest; --kernel 2 draws about 2% less board power for about 1% less hashrate. If you set --kernel 2 on an RTX 50 card for speed, remove it. Both need NVIDIA driver 580 or newer
  • RTX 30 / 40 series: one Pearl kernel. The former efficient kernel is now both faster and lower-power there, so it is the only one and --kernel no longer changes anything on these cards
  • Temperature: --temp-limit / --temp-resume follow the core temperature; the new --mem-temp-limit / --mem-temp-resume do the same for the memory on cards that report it (Windows; GDDR6X / GDDR7 boards). A card with hot memory no longer stays paused because of a core limit
  • AMD on Linux: the Volt column and voltage_mv in the API show the live core voltage, as on NVIDIA. Core and memory clocks come from the live sensors, so an idle card no longer shows its selected level
  • CMP 70HX: --mclk 5000 / 810 / 405 now holds, as on the CMP 90HX and CMP 50HX (driver 610.43.03). Pearl at 810 MHz: about 29% less board power for 0.3% less hashrate; at 405 MHz: about 32% less power for 1.3% less hashrate
  • Also: Pearl shares are sent smaller — compressed on pools that support it; a new start-up screen

ForgeMiner v2.0.1 — Quantus faster on the RTX 40, RTX 50 and RTX 30, memory clock lock on Windows for the RTX 40 and RTX 50, Pearl on the RX 9000 works on Windows again, Pearl on the RTX 20 about 2.7× faster with memory at 810 MHz, and fewer stale shares on some pools.
  • Quantus: RTX 40 +1.1%, RTX 50 +0.6%, RTX 30 +0.15%. Dev fee on Quantus, Pearl and NOID: 1%
  • Windows, RTX 40 / 50: --mclk 405 and --mclk 810 now hold, and everything is put back on exit. A high-refresh monitor on the mining card can block it — forge says so (FAIL with a hint) and keeps mining; lower the refresh rate or plug the monitor into another GPU
  • Windows: out of GPU memory while starting Pearl, the card retries with a smaller batch; each card shows which overclock settings the driver accepted and which it refused
  • Pearl on the RX 9070 XT / 9060 XT under Windows no longer stops with OpenCL error -5
  • Pearl on the RTX 20 with memory at 810 MHz (or lower): about 2.7× faster, with a new kernel for that case
  • Fewer stale shares on pools that accept a re-sent job only for a short time (GitHub issue #9)
  • AMD: Qubitcoin about +0.7% (RX 6700 XT); Conflux on the RX 7700 XT about 18% less power for 0.7% less hashrate, and on the RX 6000 the card always reaches its top speed
  • Also: --no-recover keeps a failed card stopped; the dev fee goes over TLS when your pool uses ssl:// (Kryptex), and if the dev-fee pools can't be reached the error names the cause for each address

ForgeMiner v2.0.2 — NVIDIA core voltage on Windows for the RTX 50, Pearl on the RX 9000 under Windows faster again, the pool password reaches pools that read settings from it, and SSL launchers that some networks in Russia blocked connect again.
  • Windows, RTX 50: --vddoff / --vddc now work, the same as on Linux — on top of your core lock and offset; a request the card doesn't take is put back and the card keeps mining; everything is restored on exit
  • Pearl on the RX 9070 XT / 9060 XT under Windows: a new kernel, about +13–14% over 2.0.1
  • The pool password is sent to the pool, so pools that take settings from it (payout mode, difficulty) get them
  • Some SSL launchers (Pearl, QubitCoin) now use the pool's other addresses — the same servers — that connect from Russia; a pool's TLS that gets no answer is now reported on every coin

HiveOS custom miner — upgrade with the versioned URL:
Code:
https://github.com/0xHashRaptor/ForgeMiner/releases/download/v2.0.2/ForgeMiner-2.0.2.tar.gz
The RTX 50-series (Blackwell) needs NVIDIA driver 580 or newer.

Download: https://github.com/0xHashRaptor/ForgeMiner/releases
Docker: docker pull hashraptor/forge
Pages: « 1 [2]  All
  Print  
 
Jump to:  

Powered by MySQL Powered by PHP Powered by SMF 1.1.19 | SMF © 2006-2009, Simple Machines Valid XHTML 1.0! Valid CSS!