A · cards of 24 GB or more
- RTX 3090 / 4090 / 5090 — the model fits on one card.
- Several of them? Run it once per card, each with its own directory:
TENSORCASH_HOME=/root/tcm-gpu1 /root/install_native.sh --wallet … --gpu 1 --service. The download is shared — only the first card pulls it.
B · cards under 24 GB — combined into one miner
- 2×12 GB, 4×8 GB or 8×6 GB. The model is split across the cards and they mine as a single worker. Cheap multi-GPU offers are usually the best value here.
- Run it once with no --gpu and no per-card directory — the cards are combined automatically. Do not also start one instance per card; they would fight over the same GPUs.
- Identical cards only, and 2, 4 or 8 of them (a 3-card offer uses 2). Expect roughly a sixth of a 4090 from 2×12 GB — usually well worth it at these rental prices.
Either way
- Ampere or newer. Turing (2080 Ti, Titan RTX, incl. 22 GB modded ones) has no bf16 — its shares are rejected, whatever the card count.
- CPU: AVX2 required — check the offer's CPU (old Celeron/Pentium hosts are refused by the installer)
- Instance: root access. Disk: ~35 GB free for the installer below, ~50 GB if you use our Docker image.
- Ports: none — ignore the provider's port mappings
Two ways to install — pick by what the provider lets you choose
- 1 · You can set the instance's Docker image (Clore, vast.ai): use ours. The model weights are baked in, so it mines ~2 min after boot instead of downloading them first.
24 GB+ cards only, one instance per card — it drives a single GPU and refuses anything smaller, so case B above must use option 2.
- 2 · Ubuntu image, or any card under 24 GB — run the installer inside the instance. This route handles every case, and is the only one that combines cards (case B).
One-time ~8.5 GB download (resumable — re-run to continue), then the miner auto-restarts on crashes (--service). Watch progress with tail -f /root/tensorcash-miner/miner.log; your worker appears here a few minutes after the model loads.
Why a separate installer? Rented instances put you inside a container where Docker-in-Docker is blocked, so this runs the exact same miner under proot. Tested on Clore; should work on other providers. Details in the
rental & no-Docker guide.