Google95 GB VRAM819 GB/s$4.2/hr

BGE-M3 (embed, 1024-dim) on TPU v5p (per-chip, 95GB): Does It Fit, How Fast, What It Costs

BGE-M3 (embed, 1024-dim) at FP16 needs 1.1 GB for its weights and 7.6 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
7.6 GB needed / 95 GB available
Decode speed
468.6
tok/s per stream
Aggregate
5623
tok/s @ 8 concurrent
Cost / 1M tokens
$0.2075
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 GB7.6 GB468.6562372.7 ms$0.2075✓
2× TPU190 GB7.6 GB796.71912051.5 ms$0.1220✓
4× TPU380 GB7.6 GB15937647840.8 ms$0.0610✓
8× TPU760 GB7.6 GB318730591435.4 ms$0.0305✓
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?
4K233412.3 MB✓ fits
8K116824.6 MB✓ fits
33K293.22 GB✓ fits
66K146.44 GB✓ fits
131K712.88 GB✓ fits
Other GPUs that run BGE-M3 (embed, 1024-dim)
Same workload, same quantization — ranked by value (tokens per dollar-hour).
GPUGPUsAggregate tok/sCost / 1M tok
RTX 3080 10GB1×5218$0.0106
RTX 3090 24GB1×6427$0.0108
RTX 5090 32GB1×12304$0.0113
RTX 4090 24GB1×6921$0.0161
RTX 4080 16GB1×4923$0.0169
AMD MI300X 192GB1×36391$0.0191
Frequently asked questions

Can BGE-M3 (embed, 1024-dim) run on TPU v5p (per-chip, 95GB)?

Yes. BGE-M3 (embed, 1024-dim) at FP16 needs 7.6 GB including the KV cache, and TPU v5p (per-chip, 95GB) has 95 GB — so it fits on a single GPU.

How fast is BGE-M3 (embed, 1024-dim) on TPU v5p (per-chip, 95GB)?

468.6 tokens/sec per stream and 5623 tok/s aggregate with continuous batching at 8 concurrent users. Time to first token is 72.7 ms, and inter-token latency is 2.1 ms.

What does BGE-M3 (embed, 1024-dim) on TPU v5p (per-chip, 95GB) cost per million tokens?

$0.2075 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 BGE-M3 (embed, 1024-dim) need on TPU v5p (per-chip, 95GB)?

1.1 GB for the weights at FP16, plus 6.4 GB of KV cache for 8 concurrent requests at 8K context — 7.6 GB total. Every extra concurrent user at long context costs more, so leave headroom.

Is TPU v5p (per-chip, 95GB) the right choice for BGE-M3 (embed, 1024-dim)?

It works. Compare $0.2075 per 1M tokens against the alternatives below — the cheapest option that fits the same workload is 1× RTX 3080 10GB at $0.0106, and the fastest is 1× Cerebras CS-3 (WSE-3) at 137324 tok/s.

Adjust the workload yourself

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

Open in calculator