Whisper is one of the few AI workloads where a cheap card is genuinely enough. Even large-v3 — the biggest model — needs about 10GB at FP16 and half that at INT8, so VRAM stops being the constraint almost immediately and the question becomes how fast you want the transcription to finish.
The RTX 4060 Ti 16GB is the pick for most people, running large-v3 at roughly 5-7x real-time: a one-hour recording done in about ten minutes, with enough VRAM left to keep working while it runs. Spend more only if you have a backlog measured in hundreds of hours, in which case an RTX 4090 at 12-16x real-time turns a full workday of audio into under an hour of processing.
NVIDIA GeForce RTX 4060 Ti 16GB
16GB GDDR6Fast enough for daily transcription at 5–7x real-time, with 16GB VRAM to multitask alongside Whisper large-v3. Around $425.
Affiliate links — we may earn a commission at no extra cost to you. Amazon ships globally; Shopee SG covers Singapore & ASEAN.
Who this is for
This guide covers GPU selection for running OpenAI Whisper (and faster-whisper / whisper.cpp) locally. Whether you transcribe a few files per week or process hundreds of hours in batch, the right card depends on model size and throughput requirements.
Whisper model VRAM requirements
| Model | Parameters | VRAM (FP16) | VRAM (INT8) | Relative Speed |
|---|---|---|---|---|
| tiny | 39M | ~1GB | Under 1GB | 32x real-time |
| base | 74M | ~1GB | Under 1GB | 20x real-time |
| small | 244M | ~2GB | ~1GB | 10x real-time |
| medium | 769M | ~5GB | ~3GB | 5x real-time |
| large-v3 | 1.5B | ~10GB | ~5GB | 2-3x real-time |
| turbo | 809M | ~6GB | ~3GB | 6-8x real-time |
The FP16 column is OpenAI’s own published requirement per model, not an estimate — tiny and base at ~1GB, small ~2GB, medium ~5GB, large ~10GB, turbo ~6GB. The INT8 column is roughly half that, which is what quantised runtimes like faster-whisper deliver.
Whisper is a lightweight workload compared to LLMs or image generation, but “lightweight” has a floor: large-v3 at FP16 wants 10GB, so an 8GB card is the one place where the model you pick still matters. OpenAI describes turbo as an optimised large-v3 with minimal accuracy loss, and at ~6GB it is the sensible large-class choice on 8GB.
Because Whisper runs on modest hardware, the real decision is between budget cards — compare any two directly:
| Spec | RTX 3060 12GB (used) | RTX 4060 Ti 16GB |
|---|---|---|
| VRAM | 12GB GDDR6 | 16GB GDDR6 |
| Memory bandwidth | 360 GB/s | 288 GB/s |
| SDXL (s/image) | 9.5s | 7.5s |
| Flux Dev (s/image) | 18s | 13s |
| 7B LLM (tok/s) | 20 tok/s | 25 tok/s |
| TDP | 170W | 165W |
| Street price | ~$250 used | $425 |
Transcription speed by GPU
| GPU | VRAM | large-v3 (FP16) | large-v3 (INT8) | Price |
|---|---|---|---|---|
| RTX 5090 | 32GB | ~15x real-time | ~20x real-time | ~$4,900+ |
| RTX 4090 | 24GB | ~12x real-time | ~16x real-time | ~$2,200 |
| RTX 4070 Ti Super | 16GB | ~8x real-time | ~11x real-time | ~$800 |
| RTX 4060 Ti 16GB | 16GB | ~5x real-time | ~7x real-time | ~$425 |
| RTX 4060 | 8GB | Won’t fit (needs ~10GB) | ~6x real-time | ~$479 |
| RTX 3060 12GB | 12GB | ~3x real-time | ~4x real-time | ~$250 used |
“Real-time” means a 1-hour audio file takes 1 hour. At 5x real-time, that same file finishes in 12 minutes. At 12x, it takes 5 minutes.
Best picks by use case
Casual transcription (a few files per week)
RTX 4060 (~$479) — 8GB does not hold large-v3 at FP16, which needs ~10GB, so run it at INT8 (~5GB) or use the turbo model (~6GB). Either way you get roughly 6x real-time and a 1-hour recording finishes in about ten minutes. Whisper is still the workload where 8GB stretches furthest — just not at full precision on the largest model.
Check NVIDIA GeForce RTX 4060 on Amazon→Buy on Shopee SG→Regular transcription (daily use, multiple files)
RTX 4060 Ti 16GB (~$425) — Faster compute pushes large-v3 to 5-7x real-time. The extra VRAM means you can run Whisper alongside other applications. Best value for anyone who transcribes regularly.
Check NVIDIA GeForce RTX 4060 Ti 16GB on Amazon→Buy on Shopee SG→Batch processing (hundreds of hours)
RTX 4090 (~$2,200) — At 12-16x real-time, you process a full 8-hour workday of recordings in under an hour. The raw throughput makes a real difference when you have backlogs of content to transcribe.
Check NVIDIA GeForce RTX 4090 on Amazon→Buy on Shopee SG→Optimization tips
- Use faster-whisper instead of the original OpenAI implementation — it is 2-4x faster with CTranslate2 backend
- Use INT8 quantization — minimal accuracy loss with 30-50% faster inference
- Batch process by splitting long audio into chunks and processing in parallel
- Use VAD (Voice Activity Detection) to skip silence — saves 10-30% processing time on recordings with pauses
- Run whisper.cpp for maximum CPU+GPU efficiency on lower-end hardware
Which GPU should you buy?
Transcribing a few files occasionally: The RTX 4060 at $479 handles Whisper large-v3 at 4-6x real-time. Good enough for personal use.
Daily transcription for work or content creation: The RTX 4060 Ti 16GB at ~$425 is the sweet spot. Reliable speed and enough VRAM for multitasking.
Batch processing large audio archives: The RTX 4090 at $2,200 cuts processing time to minutes per hour of audio. Worth it if transcription is a core part of your workflow.
Whisper is your only AI workload: Do not overbuy. Even a $250 used RTX 3060 runs Whisper large-v3 at 3-4x real-time.
Common mistakes to avoid
- Buying a flagship GPU just for Whisper. Whisper is one of the least demanding AI workloads. Unless you also run LLMs or generate images, a mid-range card is more than enough.
- Running the tiny or base model to save VRAM. The quality difference between tiny and large-v3 is dramatic, especially for non-English or noisy audio. Use large-v3 with INT8 if VRAM is tight.
- Using the original Whisper implementation. Switch to faster-whisper (CTranslate2) for a 2-4x speed improvement with identical accuracy.
- Processing long files as a single chunk. Split audio into 30-second segments for better GPU utilization and lower peak VRAM.
Final verdict
| Budget | GPU | Whisper Speed |
|---|---|---|
| $250 | RTX 3060 12GB (used) | 3-4x real-time |
| $479 | RTX 4060 | 4-6x real-time |
| ~$425 | RTX 4060 Ti 16GB | 5-7x real-time |
| $2,200 | RTX 4090 | 12-16x real-time |
NVIDIA GeForce RTX 4060
8GB GDDR6The sweet spot for Whisper transcription — 4–6x real-time speed on large-v3 at a price that makes sense for a single-purpose workload.
Affiliate links — we may earn a commission at no extra cost to you. Amazon ships globally; Shopee SG covers Singapore & ASEAN.
Whisper is the rare AI workload where budget GPUs shine. A $479 RTX 4060 transcribes faster than any human typist. For broader AI use beyond transcription, see our general AI GPU guide and best GPUs under $500.
Whisper runs well on almost anything with a GPU. Buy for your other AI workloads first and let Whisper ride along for free.
Common questions about GPUs for Whisper
How much VRAM does Whisper large-v3 need?
Whisper large-v3 needs roughly 10 GB of VRAM at FP16, dropping to around 5 GB with INT8 quantization at minimal accuracy cost. That makes it one of the lightest AI workloads — even 8 GB and 12 GB cards can run the largest Whisper model, so the real question is transcription speed rather than whether the model fits at all.
Do you need an expensive GPU for real-time transcription?
No. Even budget cards transcribe far faster than real-time: a used RTX 3060 12GB runs large-v3 at roughly 3-4x real-time, and an RTX 4060 reaches roughly 4-6x. Flagship GPUs like the RTX 4090 only make sense for batch processing large archives, where roughly 12-16x real-time turns a full workday of recordings into under an hour of processing.
Is 8GB of VRAM enough for Whisper?
Yes. Whisper is one of the few AI workloads where 8 GB is not a limitation — large-v3 fits in roughly 10 GB at FP16 and around 5 GB with INT8, so an 8 GB card like the RTX 4060 runs the largest model at roughly 4-6x real-time. For occasional transcription of a few files per week, that is more than fast enough.
What is the fastest way to speed up Whisper transcription?
Switch from the original OpenAI implementation to faster-whisper, which uses the CTranslate2 backend and is roughly 2-4x faster with identical accuracy. Add INT8 quantization for another noticeable speedup with minimal quality loss, and use voice activity detection to skip silence, which saves roughly 10-30% on recordings with pauses.