Combine Multiple Computers for Machine Learning

To combine multiple computers for machine learning with TensorTurn, install the one-line agent on each laptop or desktop you own and they join a single pool of workers. From there you can run one job across all of them — ensemble mode trains a diverse model on each machine and votes, or fused mode shards the data and periodically averages weights into one model. Because the agent is outbound-HTTPS-only, machines on different home and office networks combine without any port-forwarding or VPN setup.

Turning spare computers into a training cluster

Most people have more idle compute than they realize — a gaming desktop, an old workstation, a partner's laptop. TensorTurn lets you pool them without networking expertise. Each machine runs the same one-line agent, connects out over HTTPS, and shows up in your dashboard with live CPU, RAM, GPU, VRAM, and temperature readings.

• Run the install command on each computer (Windows, macOS, or Linux) — about a minute each.
• Each machine registers with a SHA-256-hashed token and appears as a worker.
• Machines can sit on completely different networks; outbound-only means no shared LAN is required.
• Start a run and choose how the machines cooperate.
• The work-stealing scheduler keeps faster machines busier and requeues work from any that drop.

One job across many machines

There are two honest ways to spread a single job. Ensemble mode has each computer independently train a full model on the whole dataset; the models then vote, which usually beats any single one and needs almost no communication. Fused mode splits the data across the computers, trains in parallel, and periodically averages the weights (DiLoCo-style) into one final model — useful when the dataset is large and you want a single deployable artifact.

What combining computers does — and does not — do

Pooling machines multiplies how many models you can train at once and how much data you can push through in parallel. It does not let a model that is too big for one computer's VRAM span several computers.

• Each computer must hold the entire model in its own memory; layers are never split across machines.
• Adding computers raises throughput and ensemble diversity, not the maximum model size.
• This is embarrassingly-parallel work or periodic weight-averaging — not tensor/pipeline parallelism or WAN all-reduce.
• It is your own pool of machines, single-tenant; you are not borrowing strangers' GPUs.
• The slowest machine can bottleneck fused sync, though work-stealing softens uneven speeds.

A realistic example

Say you connect a desktop with an RTX 3090 and two laptops. In ensemble mode all three train different models on your full tabular dataset and vote — you get an accuracy lift with essentially no network traffic. In fused mode the 3090 automatically takes a larger data shard than the laptops thanks to work-stealing, and the three periodically average weights into a single model you can deploy as an API. If one laptop closes its lid mid-run, its work is requeued to the others.