⛏ QUANTUS POOL — NATIVE MINER SETUP

Quick Start

The native miner is a patched build of the official quantus-miner with pool-mode support (--payout-address + --worker-name). Three things to do: download the binary, get a payout address, run the miner.

1. Download

quantus-miner v4.0.2 (Windows, Linux) · v3.2.0 (macOS) — patched for pool mode
⬇ Windows x86_64 · Linux x86_64 · macOS arm64 · SHA256SUMS

After download, make the binary executable (Linux/macOS):

chmod +x quantus-miner-*

xattr -d com.apple.quarantine quantus-miner-darwin-arm64

2. Get a payout address

You need a Quantus address (starts with qz…). Options:

• Phone wallet: quantus.com/wallet (iOS, Android, Web)
• CLI wallet: quantus-cli — quantus wallet create --name myrig

Your rewards will be credited to whatever address you put on the --payout-address flag. The address is sent over the wire on every connection, so you can use different addresses on different rigs.

3. Run the miner

Minimum config (CPU only, auto-detected workers):

quantus-miner serve \
  --node-addr 38.190.136.41:9890 \
  --auth-token pool \
  --tls-cert-sha256 3bda8193b9f170a3fc3d8c482c9c51cdc50706ae1a3efb1a575cf18f43763e6b \
  --payout-address qzYOUR_ADDRESS_HERE \
  --worker-name my-rig

What to expect in the log:

⛏ Sent pool-hello (payout=qz…, worker=my-rig)
⛏ Bidirectional stream established
⛏ Received job: id=…, hash=0x…
🎉 GPU worker 0 found solution! …
⛏ Job … completed: N hashes in X.XXs (Y MH/s)

Confirm on the dashboard: your worker appears on the Connected Miners table within ~5s with a non-zero hashrate. Click your address to see the per-worker miner detail page.

NVIDIA CUDA miner (fastest for GeForce / RTX)

A separate, from-scratch CUDA implementation of QPoW for NVIDIA GPUs. On an RTX 5090 it sustains ~1.2–1.3 GH/s against this pool, versus ~330 MH/s for the standard miner above — about 3.5–4× more shares for the same electricity. It speaks the pool's WebSocket protocol directly (no bridge, no auth token), and its hashes are verified bit-exact against the reference implementation (8000 + 4096 known-answer vectors, plus every share is re-checked by the pool). The standard miner remains available above for CPU mining, Intel/Apple GPUs and older NVIDIA cards; AMD RDNA3 owners should use the native AMD miner below.

Requirements

• NVIDIA GPU: GeForce RTX 20/30/40/50 series, or datacenter Turing/Ampere/Hopper/Blackwell (compute capability 7.5–12.x).
• NVIDIA driver 580 or newer (CUDA 13 runtime; no CUDA toolkit install needed, the runtime is built in).
• Windows 10/11 x64, or Linux x86_64 with glibc 2.35+ (Ubuntu 22.04, Debian 12, RHEL 9 or newer).
• Tuned and tested on the RTX 5090 (Blackwell). Other cards are supported but were not benchmarked — the bench command below tells you what yours does.

1. Download

qpow-cuda v1.0.8 (build 2026-09-14) — prints its version with version
⬇ Windows x86_64 (CUDA) · Linux x86_64 (CUDA) · SHA256SUMS

Linux: chmod +x qpow-cuda-linux-x86_64. Windows: the binary is unsigned, so SmartScreen may ask you to confirm the first launch.

2. Run

Payout address as above (qz…). One line, no other flags needed:

qpow-cuda-windows-x86_64.exe pool ws://quantusminer.com:9901 qzYOUR_ADDRESS_HERE my-rig

Linux: ./qpow-cuda-linux-x86_64 pool ws://quantusminer.com:9901 qzYOUR_ADDRESS_HERE my-rig

Full form: pool <ws_url> <payout> [worker] [submit_gap_s] [ngpu]. For several NVIDIA cards in one machine pass the card count as the last argument (each card connects as worker-gpu0, -gpu1, … and mines a different nonce range):

qpow-cuda-linux-x86_64 pool ws://quantusminer.com:9901 qzYOUR_ADDRESS_HERE my-rig 0.2 2

What to expect in the log (the rate in parentheses is the last 60 s; the first minutes read low while the pool's difficulty ramps up):

[gpu0] NVIDIA GeForce RTX 5090 (170 SMs) worker=my-rig
[gpu0] connected to ws://quantusminer.com:9901
[gpu0] HelloAck received
[gpu0] [job 1234] diff=1000 target=0041…
[gpu0] [300s] 1207.3 MH/s (60s: 1213.8) | diff=4399554609 submitted=… accepted=… rejected=… staleskip=… throttled=…

Measured hashrates (this binary, build 1.0.7 (2026-09-12))

Offline bench on stock power limits, measured 12 Sep 2026. Pool-mode rates are 1–5% lower; most cards are power-limited under this load, so cooling and power cap decide the sustained clock. Untested card? MH/s ~= SMs x sustained GHz x k, with k = 2.9 (Blackwell), 2.3 (Ada), 2.2 (Ampere), 1.65 (Turing). Blackwell rows below were re-measured on 1.0.7; other rows are 1.0.5-era and 1.0.6/1.0.7 added no gain on those architectures.

Benchmark and self-test

qpow-cuda-windows-x86_64.exe bench 10          # ~10 s offline throughput test, prints MH/s per GPU
qpow-cuda-windows-x86_64.exe mine <32-byte-hex-header> 100000000   # find one solution at a fixed difficulty

The miner auto-reconnects with backoff if the pool or your network drops, skips solutions whose job rolled over while they were on the GPU, and re-checks nothing on the CPU — the pool does that for every share. Keep the GPU cool: these cards are power-limited under this load, so a well-cooled card runs 10–15% faster than a hot one.

AMD GPU miner (ROCm / HIP — Linux, RDNA3)

A native AMD implementation of QPoW. On a Radeon RX 7900 XTX it sustains ~175 MH/s against this pool, versus ~96 MH/s for the standard miner on the same card — the standard miner has no AMD kernel and falls back to a generic Vulkan path. Hashes are verified bit-exact against the reference implementation (8000 known-answer vectors), and every candidate nonce is re-hashed through that verified path before it is submitted.

Worth knowing before buying hardware for it: QPoW is dominated by 64-bit integer multiplies, which RDNA3 emulates. At ~175 MH/s an RX 7900 XTX is still beaten by a far cheaper RTX 5060 at ~265 MH/s. This exists so AMD owners can mine competitively on hardware they already have.

Requirements

• Linux x86_64 with glibc 2.38+ (Ubuntu 24.04, Debian 13, Fedora 39 or newer) — the AMD builds still carry this limit; the next rebuild drops it to 2.35 like the CUDA build. No Windows or macOS build.
• AMD RDNA3, gfx1100 only: RX 7900 XTX / XT / GRE. Other AMD GPUs are not supported by this build.
• ROCm runtime installed. The binary is ~1 MB and links against it — it is not self-contained.
• /dev/kfd must exist (ls -l /dev/kfd), and your user must be in the render and video groups. In Docker pass --device=/dev/kfd --device=/dev/dri --group-add video --group-add render.

1. Pick the build that matches your ROCm

Check your version first — ls -d /opt/rocm* or apt list --installed | grep rocm. The two builds are not interchangeable.

The ROCm 5 build is ~34 % faster. That is a property of the ROCm runtime, not the kernel — identical source compiled against 6.4 simply runs slower. · Full documentation · SHA256SUMS

If it will not start and prints libamdhip64.so.5: cannot open shared object file, you have ROCm 6 — download the -rocm6 build instead (and vice versa for .so.6). This happens in the dynamic loader before the miner runs, so it cannot warn you itself.

2. Run

chmod +x qpow-hip-linux-x86_64-rocm5, then the same payout address as above:

./qpow-hip-linux-x86_64-rocm5 pool ws://quantusminer.com:9901 qzYOUR_ADDRESS_HERE my-rig

Check your setup first with ./qpow-hip-linux-x86_64-rocm5 version — it prints the ROCm version the build targets, the runtime it found and your GPU, and warns if they do not match.

qpow-hip 0.1.0 | built for ROCm 5.7 | runtime ROCm 5.x | device: AMD Radeon RX 7900 XTX (gfx1100)
[job 3018] diff=1000 target=0041…
[300s] 175.6 MH/s | diff=595790077 submitted=… accepted=… rejected=…

GPU Selection

Multi-GPU systems are common (e.g., a laptop with an iGPU + dGPU, a workstation with multiple cards). The miner uses wgpu, which enumerates every GPU it can see across every supported backend (Vulkan, DirectX 12, Metal), so your "RTX 5090" may appear three or four times in the list. Picking the right one matters.

Step 1: List the GPUs

quantus-miner gpu-list

Output looks like this on Windows (RTX 5090 + iGPU + Microsoft fallback):

idx | name                                     | backend  |     vendor |     device | type
  0 | AMD Radeon(TM) 890M Graphics             | Vulkan   | 0x00001002 | 0x0000150e | IntegratedGpu
  1 | NVIDIA GeForce RTX 5090                  | Vulkan   | 0x000010de | 0x00002b85 | DiscreteGpu
  2 | NVIDIA GeForce RTX 5090                  | Vulkan   | 0x000010de | 0x00002b85 | DiscreteGpu
  3 | NVIDIA GeForce RTX 5090 Laptop GPU       | Vulkan   | 0x000010de | 0x00002c58 | DiscreteGpu
  4 | NVIDIA GeForce RTX 5090 Laptop GPU       | Vulkan   | 0x000010de | 0x00002c58 | DiscreteGpu
  5 | AMD Radeon(TM) 890M Graphics             | Dx12     | 0x00001002 | 0x0000150e | IntegratedGpu
  6 | NVIDIA GeForce RTX 5090                  | Dx12     | 0x000010de | 0x00002b85 | DiscreteGpu
  7 | NVIDIA GeForce RTX 5090 Laptop GPU       | Dx12     | 0x000010de | 0x00002c58 | DiscreteGpu
  8 | Microsoft Basic Render Driver            | Dx12     | 0x00001414 | 0x0000008c | Cpu

Same physical GPU often appears two or three times under different backends. Vulkan tends to perform best on NVIDIA & AMD discrete cards on Windows and Linux; DirectX 12 is the fallback. macOS only has Metal (via Vulkan→Metal translation).

Step 2: Choose how many GPUs to use

As of miner v4, integrated GPUs are skipped automatically whenever a discrete card is present, so you no longer have to hand-pick an index to avoid the iGPU — --gpu-devices 1 lands on your first discrete card.

Step 3: Examples

Single discrete GPU (RTX 5090 on Vulkan, ignore the iGPU):

quantus-miner serve \
  --node-addr 38.190.136.41:9890 \
  --auth-token pool \
  --tls-cert-sha256 3bda8193b9f170a3fc3d8c482c9c51cdc50706ae1a3efb1a575cf18f43763e6b \
  --payout-address qz… \
  --worker-name rtx5090-rig \
  --cpu-workers 0 \
  --gpu-devices 1

Two GPUs in parallel (desktop 5090 + laptop 5090, both on Vulkan):

quantus-miner serve \
  --node-addr 38.190.136.41:9890 \
  --auth-token pool \
  --tls-cert-sha256 3bda8193b9f170a3fc3d8c482c9c51cdc50706ae1a3efb1a575cf18f43763e6b \
  --payout-address qz… \
  --worker-name dual-5090 \
  --cpu-workers 0 \
  --gpu-devices 2

CPU + GPU mixed (4 CPU threads + one dGPU):

quantus-miner serve \
  --node-addr 38.190.136.41:9890 \
  --auth-token pool \
  --tls-cert-sha256 3bda8193b9f170a3fc3d8c482c9c51cdc50706ae1a3efb1a575cf18f43763e6b \
  --payout-address qz… \
  --worker-name mixed-rig \
  --cpu-workers 4 \
  --gpu-devices 1

All CLI flags

Also: quantus-miner gpu-list (enumerate GPUs) and quantus-miner benchmark (10s engine speed test, no pool needed).

Running as a service

For a long-lived rig you'll want auto-restart on crash and on reboot. On Linux with systemd:

sudo tee /etc/systemd/system/quantus-miner.service <<'EOF'
[Unit]
Description=Quantus Native Miner (pool)
After=network-online.target
Wants=network-online.target

[Service]
Type=simple
User=mining
WorkingDirectory=/home/mining
ExecStart=/home/mining/quantus-miner-linux-x86_64 serve \
    --node-addr 38.190.136.41:9890 \
    --auth-token pool \
    --tls-cert-sha256 3bda8193b9f170a3fc3d8c482c9c51cdc50706ae1a3efb1a575cf18f43763e6b \
    --payout-address qzYOUR_ADDRESS \
    --worker-name my-rig \
    --cpu-workers 0 \
    --gpu-devices 1
Restart=on-failure
RestartSec=10
LimitNOFILE=65535

[Install]
WantedBy=multi-user.target
EOF

sudo systemctl daemon-reload
sudo systemctl enable --now quantus-miner
sudo journalctl -u quantus-miner -f

On Windows, easiest is NSSM to wrap the .exe as a service.

Troubleshooting

Connection refused / can't connect to 38.190.136.41:9890

• QUIC is UDP. Some corporate / café firewalls block outbound UDP. Try a different network.
• Verify port 9890 is reachable: on Linux, nc -zvu 38.190.136.41 9890.
• Don't use a hostname — --node-addr needs an IP:port literal.

"payout address rejected"

• Re-type the address. SS58 validation catches Cyrillic homographs (а/о vs a/o), embedded spaces, wrong-length addresses.
• Verify with the official wallet: it shows the canonical form.

Hashrate drops, then climbs again

• That's vardiff. Pool starts you at low difficulty (lots of shares, low value) and ramps you up to match your real hashrate. Stabilizes in 1-2 minutes.
• On the dashboard, your worker's Diff column will climb steadily until it matches your throughput.

Browser GPU on Mac doesn't work, native does

• Known issue. The WebGPU client on Safari/Chrome on Mac has a separate ~3% invalid-hash rate at high difficulty that the native miner doesn't have. Use the native binary on Mac for reliable hashrate.

Multiple miners on one host

• Set a different --metrics-port on each (default 9900 collides).
• Use distinct --worker-names so they show separately on the dashboard.