One RTX 5090, Qwen3-Coder-30B, measured: 9,024 tok/s whole-card, 86 tok/s per request. Serves paid inference when there is demand, mines PRL when there is not. Serving pays in USDC.
One GPU, two revenue streams: PRL mining plus paid inference.
Serving revenue settles in USDC — you keep 80%. Mining pays in PRL, with a USDC option.
An RTX 5090 serving at 9,024 tok/s breaks even at 6.9% utilisation vs its median rent. Idle time mines PRL.
Datacenter H100/H200 via the inference gateway; consumer 3090/4090/5090 via stratum.
Every card here mines; any of them can also register as a serving node.
| GPU | PRL hashrate | Source |
|---|---|---|
| NVIDIA H100 | 743 TH/s | Measured |
| NVIDIA H200 | H100-class | Estimate |
| RTX 5090 | 350 TH/s | Measured |
| RTX 4090 | 310 TH/s | Measured |
| RTX 3090 | 67 TH/s | Measured |
Measured = our own single-card runs (2026-10-03/04). The H200 row is an estimate we have not verified.
One GPU, one command, two payouts: inference in USDC, idle-time mining in PRL. Private beta — request access.
A 24 GB+ NVIDIA card on Linux · your PRL address (pool account, receives mined PRL) · your USDC address on Base (receives inference earnings) · an API key from the dashboard.
# asks for the addresses and key from step 1 curl -fsSL https://pool.linkrra.com/install.sh | sh
Downloads the model (~17 GB), registers your node, serves paid requests when there is demand — and mines PRL when there is not. Paid work always preempts mining.
GET /v1/earnings.Any supported GPU (3090 / 4090 / 5090 / RDNA2 / Arc), algorithm pearlhash — just your PRL address, no registration.
SRBMiner-MULTI --algorithm pearlhash \ --pool stratum.linkrra.com:5567 \ --wallet <your-prl-address> --worker rig1