Compute Substrate Forge

A high-performance CUDA SHA-256d miner for the Compute Substrate (CSD) network. One binary mines every NVIDIA RTX 30/40/50-series GPU in your rig, from a single process, with a live per-GPU dashboard.

• Fast: ~+5% faster kernel than the reference miner (hoisted SHA rounds + tuned occupancy).
• Multi-GPU: --all-gpus runs every card in one process, each on its own pool connection (distinct extranonce → no nonce collision, no duplicate work).
• Per-card auto-tune with a persistent cache, NVML telemetry (temp/power/ efficiency), and a configurable pool.

---

Which binary?

Binary	GPUs	Driver
compute-substrate-forge	RTX 3000 / 4000 / 5000	CUDA 12.8+ / 13.x
compute-substrate-forge-cuda124	RTX 3000 / 4000	CUDA 12.4
compute-substrate-forge-hiveos.tar.gz	3000 / 4000 / 5000 (HiveOS)	—

The main binary is universal: a compute_75 kernel that the driver JIT-compiles to each card's native code — same full speed on Blackwell as a native build. Use the cuda124 build only for rigs stuck on a CUDA 12.4 driver (no Blackwell).

---

Quick start

# 1. Cuda >13.0 Linux (Series 5000/4000/3000) mine every GPU to your address
./compute-substrate-forge --all-gpus --address YOURADDRESS --pool pool.yamaduo.no:3333  --worker WORKERNAME --auto-tune

# 2. Cuda >12.4 Linux (Serie 4000/3000) mine every GPU to your address
./compute-substrate-forge-cuda124 --all-gpus --address YOURADDRESS --pool pool.yamaduo.no:3333  --worker WORKERNAME --auto-tune

# 3. HiveOS in the shell (Series 5000/4000/3000) mine every GPU to your address
./csd-gpu-miner --all-gpus --address YOURADDRESS --pool pool.yamaduo.no:3333  --worker WORKERNAME --auto-tune

Example Flighsheet Hive-OS

Options (the ones you need)

Flag	What
--address <addr20>	required — your payout address
--all-gpus	mine every GPU in one process (recommended for a rig)
--device <n>	mine a single GPU instead (index from devices)
--pool <host:port>	mine to a different pool (default: built-in)
--worker <name>	rig name for the pool dashboard (display-only)
--auto-tune	benchmark + tune each card's launch geometry (see below)
--backend <auto|cuda|cpu>	backend select (default auto)
--stats-port <p>	xmrig-style /1/summary JSON on localhost:p
--config <file.toml>	load settings from a TOML file

Subcommands: newwallet, selftest, devices, bench. Advanced flags (geometry, CPU, watchdog, thermal, …) still exist but are hidden from --help for clarity.

Config file (config.toml, or --config):

address = "your40hexaddr20"
pool    = "pool.yamaduo.no:3333"   # optional
worker  = "rig1"                   # optional

---

Auto-tune (per card)

--auto-tune benchmarks each GPU's launch geometry and picks the fastest, per card, and persists each card's winner to ~/.config/csd-pool-miner/ autotune.toml. Run it once; later starts (without --auto-tune) reuse every card's tuned geometry automatically. Recommended on first run:

./compute-substrate-forge --all-gpus --address YOURADDRESS --pool pool.yamaduo.no:3333  --worker WORKERNAME --auto-tune

---

HiveOS

Use compute-substrate-forge-hiveos.tar.gz as a custom miner. Full turnkey flight-sheet steps are in FLIGHTSHEET.md inside the tarball. In short:

1. Host the tarball at a direct URL (GitHub release, web server, …).
2. Flight sheet → Miner Custom → Installation URL = that URL, Wallet template = %WAL%, Pool URL = pool.yamaduo.no:3333 (or empty), Extra config arguments = --all-gpus.
3. Apply to your rigs — HiveOS installs and runs it; stats show in the UI.

---

Live dashboard

--all-gpus prints a refreshing per-GPU table (hashrate, temp, power, efficiency, accepted/rejected shares per card + a rig total). Under it only new jobs and accepted/rejected shares scroll. NVML telemetry (temp/power) is built in.

 GPU  Device                 Hashrate      Temp     Power     Shares
  0   RTX 5080               7.16 GH/s    63 C    224 W    a=3 r=0  eff 31.9 MH/s/W
  ...
 TOTAL                       30.56 GH/s           1207 W    a=13 r=0

---

Performance notes

• SHA-256d is current-limited (EDP) on Blackwell: cards run below their power limit (e.g. ~270 W of a 360 W 5080) at ~3.0 GHz core — this is normal and efficient. Raising the power limit does not help; only a V/f-curve undervolt (needs nvidia-settings + X11) can push further.
• A few low difficulty share rejects at startup are normal (pool vardiff ramping).

---

Voluntary donations (separate, optional): 0x91723defb8e9dac838f0faf2ec6beec3282d1c3c.

License

See LICENSE / LICENSE-APACHE / LICENSE-MIT and TRADEMARK.md. This is opt-in mining software you run on your own hardware; it mines only when you start it, to your own payout address.