Google95 GB VRAM819 GB/s$4.2/hr

Pixtral 12B (VLM) on TPU v5p (per-chip, 95GB): Does It Fit, How Fast, What It Costs

Pixtral 12B (VLM) at FP16 needs 24.0 GB for its weights and 34.7 GB in total at 8K context with 8 concurrent users. That means it fits on a single TPU v5p (per-chip, 95GB). Every figure is computed by tokcalc's open-source capacity model.

Fits?
Yes, 1 GPU
34.7 GB needed / 95 GB available
Decode speed
22.18
tok/s per stream
Aggregate
266.2
tok/s @ 8 concurrent
Cost / 1M tokens
$4.38
1× TPU v5p (per-chip, 95GB)
What changes as you add TPU v5p (per-chip, 95GB)s
Tensor-parallel degree vs VRAM, throughput, and cost. Cost scales linearly with GPU count; decode throughput scales near-linearly too, so cost per token stays roughly flat.
GPUsTotal VRAMVRAM neededDecode tok/sAggregate tok/sTTFTCost / 1M tokFits?
1× TPU95 GB34.7 GB22.18266.2931.8 ms$4.38✓
2× TPU190 GB34.7 GB37.71905.0484.9 ms$2.58✓
4× TPU380 GB34.7 GB75.423620257.4 ms$1.29✓
8× TPU760 GB34.7 GB150.814480143.7 ms$0.6446✓
Concurrency vs context length
On 1× TPU v5p (per-chip, 95GB) at FP16. KV cache scales linearly with context, so this is the number of users your GPUs can actually serve.
ContextConcurrent usersKV per requestFits?
4K105687.2 MB✓ fits
8K521.34 GB✓ fits
33K135.37 GB✓ fits
66K610.74 GB✓ fits
131K321.47 GB✓ fits
Other GPUs that run Pixtral 12B (VLM)
Same workload, same quantization — ranked by value (tokens per dollar-hour).
GPUGPUsAggregate tok/sCost / 1M tok
RTX 5090 32GB2×1980$0.1403
AMD MI300X 192GB1×1723$0.4032
A100 40GB SXM41×503.8$0.4963
AMD MI325X 288GB1×1950$0.4986
A100 80GB SXM41×650.0$0.5128
Intel Gaudi 3 128GB1×1073$0.5180
Frequently asked questions

Can Pixtral 12B (VLM) run on TPU v5p (per-chip, 95GB)?

Yes. Pixtral 12B (VLM) at FP16 needs 34.7 GB including the KV cache, and TPU v5p (per-chip, 95GB) has 95 GB — so it fits on a single GPU.

How fast is Pixtral 12B (VLM) on TPU v5p (per-chip, 95GB)?

22.18 tokens/sec per stream and 266.2 tok/s aggregate with continuous batching at 8 concurrent users. Time to first token is 931.8 ms, and inter-token latency is 45.1 ms.

What does Pixtral 12B (VLM) on TPU v5p (per-chip, 95GB) cost per million tokens?

$4.38 per 1M output tokens at 1× TPU v5p (per-chip, 95GB) ($4.20/hr). That is the marginal hardware cost — add idle capacity, ops, and egress for the real figure.

How much VRAM does Pixtral 12B (VLM) need on TPU v5p (per-chip, 95GB)?

24.0 GB for the weights at FP16, plus 10.7 GB of KV cache for 8 concurrent requests at 8K context — 34.7 GB total. Every extra concurrent user at long context costs more, so leave headroom.

Is TPU v5p (per-chip, 95GB) the right choice for Pixtral 12B (VLM)?

It works. Compare $4.38 per 1M tokens against the alternatives below — the cheapest option that fits the same workload is 4× RTX 3080 10GB at $0.0662, and the fastest is 1× Cerebras CS-3 (WSE-3) at 6500 tok/s.

Adjust the workload yourself

Same formulas, your context length, concurrency, and quantisation.

Open in calculator