How much hardware does Whisper need? Model sizes, memory, and what a GPU or a CPU can do, from OpenAI’s table and a public benchmark.
Short answer. Whisper needs far less hardware than a chat model. OpenAI’s table lists about 10 GB of VRAM for the large model, about 6 GB for the faster turbo model and about 2 GB for the small model. Optimised runtimes use less: in the faster-whisper project’s test the large-v2 model needed 2.9 to 6.1 GB of GPU memory on an 8 GB RTX 3070 Ti and transcribed 13 minutes of audio in 16 to 63 seconds. A CPU can run smaller models too. Accuracy varies widely by language, so test your own Hindi and Hinglish recordings before choosing a size.
01
OpenAI’s table of sizes
The README says relative speeds were measured on an A100 and the real-world speed varies with the language, the speaking speed and the hardware. It describes turbo as an optimised version of large-v3 with minimal loss of accuracy; turbo is not trained for translation.
Size
Parameters
Required VRAM
Relative speed (A100, English)
tiny
39 million
about 1 GB
about 10 times
base
74 million
about 1 GB
about 7 times
small
244 million
about 2 GB
about 4 times
medium
769 million
about 5 GB
about 2 times
large
1,550 million
about 10 GB
1 time
turbo
809 million
about 6 GB
about 8 times
02
What optimised runtimes actually used
These are the faster-whisper project’s own measurements: the GPU rows on an NVIDIA RTX 3070 Ti with 8 GB (CUDA 12.4), the CPU rows on an Intel Core i7-12700K with 8 threads. The large model on a CPU was not part of the test. Plan with OpenAI’s larger numbers if you do not yet know which runtime you will use.
Setup (13 minutes of audio, beam size 5)
Memory
Time
Large-v2 on GPU, openai/whisper, 16-bit
4,708 MB of VRAM
2 min 23 s
Large-v2 on GPU, whisper.cpp with flash attention, 16-bit
4,127 MB of VRAM
1 min 05 s
Large-v2 on GPU, faster-whisper, 16-bit
4,525 MB of VRAM
1 min 03 s
Large-v2 on GPU, faster-whisper, 8-bit
2,926 MB of VRAM
59 s
Large-v2 on GPU, faster-whisper, 16-bit, batch size 8
6,090 MB of VRAM
17 s
Small on CPU, faster-whisper, 8-bit
1,477 MB of RAM
1 min 42 s
Small on CPU, whisper.cpp, 32-bit
1,049 MB of RAM
2 min 05 s
03
What to buy
For occasional use, a modern CPU with 8 GB or more of memory runs the smaller models, more slowly than a GPU. The next page covers the CPU case.
For regular meeting transcription, a GPU with 6 to 8 GB free is enough for large or turbo. The large-v3 weights are a 3.1 GB file at 16-bit, so the model is small next to a chat model.
On a machine that also runs a language model, count Whisper’s few gigabytes in the memory budget; the language model’s cache usually needs far more.
faster-whisper’s GPU path needs NVIDIA’s cuBLAS and cuDNN 9 libraries for CUDA 12, so the NVIDIA route is the most direct for that runtime.
04
Language matters more than hardware
OpenAI’s README says Whisper’s performance varies widely depending on the language and publishes error rates by language for large-v3 and large-v2. Hardware will not fix a model that is weak in your language or accent, so test real recordings. ANVE’s meeting minutes work from transcripts in Hindi, English and Hinglish, with every action item quoting the transcript line it came from. ANVE has not published speed figures for its launch hardware.
05
Questions people ask
Does Whisper need an NVIDIA GPU?
No. It runs on CPUs, and the faster-whisper project’s benchmark includes CPU runs for both faster-whisper and whisper.cpp. The faster-whisper GPU route needs NVIDIA libraries.
How much disk space does Whisper large-v3 need?
The 16-bit weights are a file of about 3.1 GB on Hugging Face; the turbo model is about 1.6 GB. Allow extra space for the audio you are transcribing.
Is the turbo model as accurate as large-v3?
OpenAI describes turbo as an optimised large-v3 with a minimal loss of accuracy and about eight times the speed in its table, but it is not trained for translation. Test your own recordings.
06
What is not real yet
No sealed ANVE unit exists; every image of the appliance on this site is a labelled render. The software runs today on a development machine. No customers yet, no certifications, no delivery date, and the final price is not announced.
Reservations are ₹999 in India or US$12 elsewhere, refundable in full until you confirm an order in writing (terms). Offices that cannot use cloud AI can apply to be one of ten design partners.
