What is the best GPU for Whisper?

No single GPU is the best for Whisper. The right card is the one with enough video memory for the model you want to run. The turbo model needs about 6 GB of video memory ¹, and the large model needs about 10 GB ². Pick a card that clears the figure for your model, then check that your software can use it. On a Mac, the answer is different, because Apple silicon does not use a separate graphics card in the usual sense.

Video memory sets the model you can run

The official package's README gives a memory figure for each model, which is the most useful hardware requirement to start from. These are OpenAI's figures, for the official package.

By these figures, an 8 GB card has room for turbo but not for large. A card with less than 6 GB should run medium at most. These numbers describe the model's memory need, not the speed on your card, so measure the run time on a short file before you commit to a model.

NVIDIA cards are the standard GPU route

The openai-whisper command line runs on an NVIDIA GPU, through CUDA, only when PyTorch can detect one. Otherwise it falls back to the CPU ⁶. That makes an NVIDIA card with a CUDA-enabled PyTorch the most direct GPU setup for the official package.

The faster-whisper implementation has its own GPU route. Running it on a GPU requires NVIDIA libraries to be installed first ⁷. Its README also reports memory with int8 precision, which used 2926 MB of GPU memory with the large-v2 model ⁸. That is well below the 10 GB figure for large in OpenAI's table. The two numbers come from different tests and different models, so treat the lower figure as a sign that quantized runs need less memory, not as a guarantee for your card.

Apple silicon Macs use whisper.cpp or MLX, not a graphics card

A Mac with Apple silicon has no NVIDIA card, so the CUDA route does not apply. For Apple silicon, the whisper.cpp README describes the platform as a first-class target and lists Metal among its optimizations ⁹. MLX, Apple's framework for the same chips, keeps its arrays in shared memory, so operations can run on the CPU or the GPU without copying the data ¹⁰. The guide to running Whisper locally gives the install steps for both routes.

On a Mac, the memory figure that matters is the memory of the whole machine, not a separate card's video memory. Check how much memory your Mac has before you choose a model, and use the table above as a guide to what each model needs.

Which card to buy depends on the model you need

• Less than 6 GB of video memory: use medium at most, since it needs about 5 GB.
• From 6 GB to under 10 GB: turbo fits by the table's figures. Large needs about 10 GB and may not.
• 10 GB or more: the large model fits by the table's figures.
• No suitable GPU at all: Whisper still runs on the processor, more slowly. Our answer on running Whisper without a GPU covers that route.

What to do

1. Check the video memory your graphics card reports, or the memory of your Mac.
2. Start with turbo if it fits. If it does not, step down to medium.
3. Time a short file. If the speed is acceptable, keep the setup. If not, compare the other implementations in our answer on whisper.cpp versus faster-whisper.

For the install steps, see how to install Whisper AI. If the GPU is not being used at all, our answer on why Whisper is not using your GPU walks through the checks.