How to mine PEARL on a single GPU — full setup guide

PEARL mining looks weird at first glance. There's no hash race, no ASIC arms race, no SHA-256 — instead, a GPU runs large language model inference, and the model's outputs get committed to the chain as proof of work. The "useful" part of Proof-of-Useful-Work means that the same compute cycles that secure the network also produce something other people pay for: AI text completions.

This guide walks you through the minimum viable setup for mining PEARL on a single GPU — what hardware you need, what runs on it, and the math you should do before turning anything on.

Hardware requirements

The Pearl miner stack uses vLLM as the inference backend. That sets the floor for what counts as a viable card:

• Minimum: NVIDIA H100 (80 GB SXM or PCIe). Anything older than Hopper struggles with the kernel layers Pearl uses.
• Recommended: H200 (141 GB) — more headroom for larger models, slightly higher hash equivalent.
• Not supported: consumer cards (RTX 4090 and below). They lack the SM 9.0 arch features the Pearl kernel relies on.
• RAM: 64 GB system memory is comfortable; 32 GB tight but workable.
• Disk: 200 GB SSD for model weights + chain state. NVMe makes block submission noticeably snappier.
• Network: 100 Mbps symmetric is plenty. The chain itself is light; bursts come when you submit a winning block.

If you don't own an H100, rent one. Cloud GPU rentals run $1.50–$2.50 per hour depending on provider, and several have signup credit:

• RunPod gives $5–10 free credit on signup (covers 2–4 hours of testing).
• Vast.ai is the cheapest community marketplace — 30–50% under big-cloud prices for the same hardware.

Software stack — what runs where

Three processes sit on the box:

1. pearld — the Pearl full node. Listens on port 44107 for chain RPC. Maintains the local copy of the chain.
2. vLLM server — runs the inference model on the GPU. Listens on a local port and serves token completions.
3. Pearl gateway — the glue. Pulls block templates from pearld, submits inference jobs to vLLM, packages the result as a candidate block, and broadcasts winning blocks back to the network.

The gateway is what mines. The other two are dependencies it talks to.

Setup — bare-metal commands

Assuming Ubuntu 22.04 with NVIDIA drivers + CUDA 12 already installed:

# Pull the Pearl release tarball
curl -LO https://github.com/PearlChain/pearl/releases/latest/download/pearl-linux-x64.tar.gz
tar xzf pearl-linux-x64.tar.gz -C /opt
cd /opt/pearl

# Start the full node — first sync takes ~2 hours
./bin/pearld --daemon

# Install vLLM (Python 3.11 venv recommended)
python3.11 -m venv /opt/vllm-venv
/opt/vllm-venv/bin/pip install vllm
/opt/vllm-venv/bin/python -m vllm.entrypoints.openai.api_server \
    --model meta-llama/Llama-3.1-8B \
    --port 8000 &

# Start the gateway, pointing at your wallet
./bin/pearl-gateway \
    --node-rpc http://localhost:44107 \
    --vllm-url http://localhost:8000 \
    --wallet prl1pyourwalletaddresshere

Within a minute you should see the gateway log lines like template fetched height=42500 followed by share submitted and (eventually) BLOCK FOUND.

Monitoring your hashrate

The Pearl gateway prints local metrics every 30 seconds. Watch for two numbers:

• completions_served — how many inference jobs your GPU finished. This is your "hash" rate.
• share_events — how many of those completions cleared the difficulty target and counted as a share. Higher is better.

To see how you compare to the rest of the network, watch your wallet on the Lord Of Pearls dashboard — the rich list, top-miner board, and your address's drill-down page all update within seconds of a successful block.

Profitability math (do this BEFORE you plug in)

The numbers move daily, but the framework doesn't:

1. Look up the current block reward on the dashboard. As of writing it's around 2,845 PEARL per block.
2. Note the network hashrate. Your GPU produces your hashrate. The ratio of (your hashrate) / (network hashrate) is your expected share of blocks per day.
3. The chain produces ~144 blocks per hour, so 3,456 blocks per day. Multiply your share by 3,456 × 2,845 PEARL to get expected daily PEARL.
4. Multiply by current PEARL price (in USD) to get gross daily revenue.
5. Subtract GPU cost: rented H100 at $2/hr × 24 = $48/day. Plus electricity if you own.

If gross > rental cost, you're profitable. If not, wait — the network hashrate fluctuates and so does PEARL price. Single-GPU mining is profitable in some months and not in others.

Common issues

• "Templates fetched 0 in 5 min" — pearld isn't fully synced yet. Wait for sync to finish.
• "vLLM 503" — the model didn't load fully. Check vLLM logs; usually OOM, drop to a smaller model.
• "Submit rejected" — your gateway is on a fork, or the share is stale. Restart the gateway to re-sync.
• "Orphan block!" — you mined a valid block that lost the chain race to another miner. See the welcome page for the orphan tracker that tells you exactly when this happens to you.

Next steps

Once you're mining, watch the dashboard. Track your wallet's rank, your hashrate share, your orphan rate. The most profitable miners run multiple GPUs and tune their gateway settings — but the single-GPU setup above gets you on-chain in a couple of hours.

If you have questions, the PEARL Discord has an active mining channel and the gateway maintainers hang out there.