The honest answer is that Apple Silicon is excellent at one thing and mediocre at another, and which one you care about decides everything. We run a farm of Macs generating video, so this comes from operating them rather than benchmarking them.
Where it genuinely wins: memory
Unified memory means the GPU can address the whole machine. A 64GB Mac holds a model that a 24GB discrete card has to offload — and offloading is where nominally faster hardware quietly becomes slower, because you have traded compute for a bus.
That is not a small advantage. It is the difference between running a model and not running it. One of our 1080×1920 renders peaked at 44.1GB; there is no consumer NVIDIA card that holds that without splitting the work up.
Where it loses: raw throughput
Not close. A current CUDA card is substantially faster per job, and the gap is not something a software update closes. A hero clip that takes tens of minutes on an M4 is not going to take tens of seconds next year.
There is also no CUDA ecosystem. You are on MLX, which means the specific port of the model you want matters more than the silicon does. A model nobody has ported runs at zero tokens per second on the best Mac ever made.
The question that actually decides it
Is your work latency-shaped or batch-shaped?
Latency-shaped — you need one result, now, and you are waiting on it. A Mac is a poor choice. Rent the GPU, pay the meter, get your answer.
Batch-shaped — you need forty results by morning and you are asleep for most of it. A Mac is an excellent choice, and four of them are four times as excellent, because the jobs are independent and the queue is the only thing they share. Per-clip speed stops being the number that matters.
What about the Neural Engine?
Reasonable question, and the answer disappoints people. The Neural Engine is built for small, fixed, low-power inference — the on-device models Apple ships. The heavy generative work runs on the GPU via Metal, which is what MLX targets. For this kind of workload the Neural Engine is largely not the part doing the lifting, and no amount of configuration changes that.
Is a MacBook Air good for this?
For learning, yes. For a render node, the limit is memory and sustained thermals, not the chip's peak. A fanless Air throttles under a load that runs for half an hour, and the base memory configuration rules out the larger models entirely. If you are choosing deliberately, memory is the specification to spend on — more than cores, more than generation.
The verdict
Apple Silicon is a bad GPU and a good farm node: large memory, low power, already paid for, and idle every night. If that matches the shape of your work, the hardware is already in the building. The full CUDA comparison is here.