Energy use of Mac inference on MetalCloud

For some jobs, a Mac you already own uses less electricity than the rack of GPUs that would otherwise run them. It depends on the model, and on that Mac already being there.

A Mac you already have

Most of a Mac’s lifetime emissions come from making it, not from plugging it in. Apple’s product environmental reports show that. If the Mac is one your studio, agency, or team already owns, and the job runs while it would otherwise be off, you are not asking anyone to build another GPU for that work.

Buying a Mac in order to rent it out does not carry that benefit. Hosting is a way to earn from a machine you bought for your own work.

One Mac instead of several GPUs

Some models do not fit in the memory of one data-centre GPU. They are spread across several GPUs, each with its own board, plus a host computer and cooling. For those jobs, compare the electricity the Mac uses for the whole job with the electricity that full server uses for the same job: same model, same precision, same amount of output.

The wattage printed on the chip is not that comparison. For smaller models that fit on one GPU, the faster GPU can finish sooner and use less electricity per token. An H100 is ahead on Llama 7B and on quantized Llama 70B, as the Apple silicon comparison shows.

You pay for the time the job runs

You are not keeping a high-power GPU switched on between requests. If you host, a Mac left on with no job still uses electricity. Put it to sleep when it has nothing to run.

iOS builds

iOS builds have to run on a Mac. Using a Mac you already have, overnight, can stand in for a second Mac you would otherwise leave on for builds. This is not a comparison with an NVIDIA GPU.

When a GPU server can use less

You are training, or fine-tuning for speed.
The model already fits on one faster GPU.
You would be buying the Mac for this.
The Mac stays on with no work.
The other data centre buys renewable electricity. Fewer watts on your side do not automatically mean less carbon.

If you need a figure for a report

Ask for watt-hours for a named job. A single multiplier would hide which model you ran and where the machine was plugged in.

What you need	What it means
The same job on both sides	Same model, precision, prompt, and length of output. The first job is Llama 70B at FP16. The walkthrough says what has to match.
The whole machine	A meter at the plug, from the start of the job to the last token. On the GPU side, every card the model needs, plus the computer they sit in. Idle time is counted on its own, so a machine waiting for work is not mixed into the job.
Cooling, if you include it	The data hall’s cooling is its own line. It is not added quietly to the Mac, and it is not left off the GPU server.
The figure you can use	Watt-hours for that job. Not the rated watts on the chip, and not one number for every Mac job.
Carbon, if you need it	Multiply those watt-hours by the grid factor for where the Mac is plugged in, and date it. A data centre on a renewable contract can come out lower even when the Mac used fewer watt-hours.

Sources

UK Green Claims Code, from the Competition and Markets Authority, on comparisons that have to stay fair across the life of the machine.
Apple’s product environmental reports. For Macs, most lifetime greenhouse gas emissions are from making the machine.

People use MetalCloud for the memory and the price. Electricity is the smaller part of that decision, and it favours the Mac only in the cases above.