What GPU do I need to run ibm-granite/granite-speech-3.2-8b?
8.5B parameters, published in BF16. View on Hugging Face
Set up granite-speech-3.2-8bgranite-speech-3.2-8b is published by ibm-granite on Hugging Face, with 62,584 downloads and 88 likes to date. It's a GraniteSpeechForConditionalGeneration model built for automatic-speech-recognition, published natively in BF16.
VRAM required & cheapest live GPU fit
Required VRAM = weight size at each precision, plus a fixed overhead for activations and allocator fragmentation. Speech models don't build the same growing KV-cache a text model does. Memory scales primarily with input audio length. Full formula and assumptions: methodology.
| Precision | Weight size | Required VRAM | Cheapest live fit | GPUs needed | Est. $/hr (full fit) |
|---|---|---|---|---|---|
| BF16 | 15.8 GB | 19.0 GB | RTX A5000 | 1 | $0.176/hr |
| FP8 (quantized) | 7.9 GB | 9.5 GB | RTX 4070 Super | 1 | $0.121/hr |
| INT4 (quantized) | 4.0 GB | 4.7 GB | RTX 4070 Super | 1 | $0.121/hr |
A GPU is only matched to a row if its hardware supports that precision, and the primary recommendation is always a single-GPU fit when one exists.
INT4 requires a quantized checkpoint actually published for this model, check its Hugging Face page before relying on this row.
Cheapest way to run granite-speech-3.2-8b at its published (BF16) precision: 1× RTX A5000, at $0.176/hr per GPU ($0.176/hr total). Quantizing to FP8 or INT4 (rows above) can cost less, but requires a compatible quantized checkpoint to exist for this model.
granite-speech-3.2-8b: KV cache by context length
The KV cache is the memory the attention layers hold for every token of context, on top of the weights. It is not part of the flat 1.2x overhead in the table above, grows with context length and with every concurrent request, and is why a long-context deployment needs more VRAM than the table shows.
Grouped-query attention: every layer caches keys and values for a few shared KV heads. Cached per token: 163,840 bytes at 16-bit.
| Context | KV cache, one sequence |
|---|---|
| 4K | 0.63 GB |
| 32K | 5.0 GB |
| 128K | 20.0 GB |
Computed from the layer, head and window counts in the model's own configuration (ibm-granite/granite-speech-3.2-8b), at 16 bits per cached value. Each concurrent sequence needs its own cache, on top of the weights.
granite-speech-3.2-8b: common questions
Does granite-speech-3.2-8b fit on a 24 GB GPU?
Yes. At BF16 it needs 19.0 GB of VRAM, so a 24 GB card holds it with 5.0 GB to spare. A 16 GB card is not enough for it at BF16.
What is the least VRAM granite-speech-3.2-8b can run in?
4.7 GB, at INT4 (quantized), which fits a 6 GB card, against 19.0 GB at BF16. That row assumes an INT4 checkpoint has actually been published for this model, so check its Hugging Face page before planning around it.
Does quantizing granite-speech-3.2-8b lower the GPU bill?
Yes. At BF16 the cheapest live fit is one RTX A5000 at $0.176/hr. At FP8 (quantized) it drops to one RTX 4070 Super at $0.121/hr, provided a quantized checkpoint exists for it.
Weight-to-VRAM math, the fit rules, and how live prices are normalized: full methodology.
More Granite Speech models
- granite-speech-3.3-8b (8.6B, BF16)
- granite-speech-3.3-2b (3.0B, BF16)
- granite-speech-4.1-2b (2.3B, BF16)
- granite-speech-4.1-2b-plus (2.1B, BF16)
- granite-speech-5.0-470m-turboctc (473M, BF16)
Lineage
Based on granite-3.2-8b-instruct.
Alternatives at this size
Other models for automatic-speech-recognition within about a third of granite-speech-3.2-8b's 8.5B parameters, from other model lines.
- VibeVoice-ASR (8.7B, BF16)
- higgs-audio-v3-8b-stt-v2 (8.9B, BF16)
More on granite-speech-3.2-8b
Fits on a 24 GB GPU at BF16: every model that fits in 24 GB.
Fits on a 12 GB GPU at FP8: every model that fits in 12 GB.
Fits on an 8 GB GPU at INT4: every model that fits in 8 GB.
Best speech-to-text models: how granite-speech-3.2-8b ranks against the rest.
Related reading: RTX A5000 pricing and specs, Whisper variants compared, and What AI inference is.
