hashraptor/forge repository overview

⁠ForgeMiner

Native NVIDIA GPU miner. One binary, eleven coins, no dependencies.

Website⁠ · GitHub⁠ · Telegram⁠ · Discord⁠

The image is a single self-contained binary on debian:bookworm-slim — about 70 MB to pull, no CUDA toolkit required. It needs only the host NVIDIA driver, exposed through the NVIDIA Container Toolkit.

⁠Run it

docker run --rm --gpus all hashraptor/forge \
  --algorithm pearlhash \
  --wallet YOUR_PRL_WALLET \
  --worker rig01 \
  --pool prl.kryptex.network:7048

That is the whole setup. Swap the algorithm, wallet and pool for another coin.

⁠Configure with environment variables

For GPU rental platforms (Vast.ai, RunPod, TensorDock) that only let you set environment variables:

docker run --rm --gpus all \
  -e FORGE_ALGO=pearlhash \
  -e FORGE_POOL=prl.kryptex.network:7048 \
  -e FORGE_WALLET=YOUR_PRL_WALLET \
  -e FORGE_WORKER=rig01 \
  hashraptor/forge

Every command-line flag has a FORGE_* twin, so anything below works either way.

⁠Coins and pools

The dev fee is interleaved into the stream rather than mined in bursts, so your pool graph has no dips, and it is verifiable on the pool. There is no hidden second fee.

Conflux is the one coin with a memory floor worth checking before you rent: it needs roughly 9.4 GB of free video memory (an 8.7 GiB dataset), and that grows slowly as the chain advances, so an 8 GB card cannot mine it and 10 GB is tight. forge reports the actual numbers at startup rather than failing partway through.

⁠Dashboard

docker run --rm --gpus all -p 7777:7777 hashraptor/forge \
  --algorithm pearlhash --wallet YOUR_PRL_WALLET --worker rig01 \
  --pool prl.kryptex.network:7048 --api-bind 0.0.0.0:7777

Then open http://<host>:7777 for live per-GPU hashrate, temperatures, power and share history.

⁠GPUs

Every generation from Pascal (GTX 10-series) through Blackwell (RTX 50-series), plus the CMP mining cards. BTX needs Volta (V100) or Turing (RTX 20) and newer — it does not run on Pascal or A100-class cards (CMP 170HX included). Conflux runs from Pascal up on cards with more than 8 GB (GTX 1080 Ti, V100, RTX 20 and newer), not on A100-class cards. Quantus, NOID and QubitCoin run from Pascal (GTX 10-series) up, V100 included, but not on A100-class cards; a card without a kernel for the chosen coin says so in one line while the rest of the rig mines. Hopper (H100/H200) runs Pearl, Cryptix and QubitCoin; Xelis, BTX, Conflux, NOID and Quantus have no Hopper build yet. The correct kernel is picked per card at startup, so mixed rigs work and each card runs its own best settings.

Requires the host NVIDIA driver (580 or newer for the RTX 50-series) and the NVIDIA Container Toolkit⁠ for --gpus all.

⁠Tags

latest always points at the newest release. Version tags such as 1.8.8 pin that release — use one if you want a build that does not move under you.

docker pull hashraptor/forge:latest
docker pull hashraptor/forge:1.8.8

⁠Overclocking

Clocks, memory offsets, power limit and fans can be driven from the miner itself — FORGE_CCLK, FORGE_MOFF, FORGE_PLIMIT, FORGE_FAN. Each takes one value for every card, or a comma list in FORGE_GPU order; put * to leave a card untouched — e.g. -e FORGE_CCLK=1500,1500,*,1500. Inside a container this needs --privileged and the NVML device nodes; on most rental platforms it is simpler to set clocks on the host.

⁠Reliability

Set -e FORGE_TEMP_LIMIT=80 -e FORGE_TEMP_RESUME=70 (or --temp-limit/--temp-resume) and a card pauses itself automatically if it overheats, then resumes once it cools back down — useful when you don't have your own external temperature watchdog on a rental box. FORGE_OC_DELAY (or --oc-delay) delays applying your overclock by N seconds after startup, for platforms where clocking instantly on boot causes instability.

⁠CMP hardware unlock (host-side)

CMP 40HX, 50HX, 70HX and 90HX ship with a hardware throttle that keeps their hashrate at a small fraction of the chip's real capacity. Forge can install the unlock helper itself — but it's a kernel-side change on the host (helper .ko modules + depmod override + reboot), so it has to be installed on the host, not inside the container.

The right way if you're on Docker with CMP cards:

# On the host (one time, root):
curl -L -O https://github.com/0xHashRaptor/ForgeMiner/releases/latest/download/ForgeMiner.tar.gz
tar -xzf ForgeMiner.tar.gz
sudo ./ForgeMiner/*/forge --cmp-install   # builds the helper for your running kernel, reboots the host after 10 seconds
# After the host comes back up, run this Docker image as usual —
# the container sees all unlocked CMP cards at full speed.

No pinned kernel version required — the helper builds from source for whatever kernel the host is running. (NVIDIA driver 610.43.03 is still required.) CMP 30HX is detected but never touched (a permanent hardware limitation on that chip; no unlock exists for it). Full walkthrough (verification, rollback): wiki · CMP hardware unlock⁠ (Русский⁠).

⁠Log capture (rental consoles, docker logs, log files)

forge auto-detects when stdout isn't a live terminal (web-console, docker logs, cron, redirect) and switches to ASCII-only output — no ANSI color escapes, no Unicode box-drawing. Interactive docker run -it still gets the full dashboard. Force plain output explicitly with -e NO_COLOR=1 or --no-color.

Full flag reference: forgeminer.org⁠ · Releases and checksums: GitHub⁠