Google95 GB VRAM819 GB/s$4.2/hr

GPT-NeoX 20B on TPU v5p (per-chip, 95GB): Does It Fit, How Fast, What It Costs

GPT-NeoX 20B at FP8 (E4M3) needs 20.0 GB for its weights and 90.9 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
90.9 GB needed / 95 GB available
Decode speed
39.93
tok/s per stream
Aggregate
479.1
tok/s @ 8 concurrent
Cost / 1M tokens
$2.44
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 GB90.9 GB39.93479.11.48 s$2.44✓
2× TPU190 GB90.9 GB67.871629758.6 ms$1.43✓
4× TPU380 GB90.9 GB135.76516394.3 ms$0.7162✓
8× TPU760 GB90.9 GB271.526064212.2 ms$0.3581✓
Concurrency vs context length
On 1× TPU v5p (per-chip, 95GB) at FP8 (E4M3). KV cache scales linearly with context, so this is the number of users your GPUs can actually serve.
ContextConcurrent usersKV per requestFits?
4K164.43 GB✓ fits
8K88.86 GB✓ fits
33K235.43 GB✓ fits
66K170.87 GB✓ fits
131K0141.73 GB✗ does not fit
Other GPUs that run GPT-NeoX 20B
Same workload, same quantization — ranked by value (tokens per dollar-hour).
GPUGPUsAggregate tok/sCost / 1M tok
RTX 5090 32GB4×14257$0.0390
A100 80GB SXM42×3978$0.1676
H100 SXM5 80GB2×6663$0.2084
AMD MI300X 192GB1×3101$0.2240
L40S 48GB2×1718$0.2263
AMD MI325X 288GB1×3510$0.2770
Frequently asked questions

Can GPT-NeoX 20B run on TPU v5p (per-chip, 95GB)?

Yes. GPT-NeoX 20B at FP8 (E4M3) needs 90.9 GB including the KV cache, and TPU v5p (per-chip, 95GB) has 95 GB — so it fits on a single GPU.

How fast is GPT-NeoX 20B on TPU v5p (per-chip, 95GB)?

39.93 tokens/sec per stream and 479.1 tok/s aggregate with continuous batching at 8 concurrent users. Time to first token is 1.48 s, and inter-token latency is 25.0 ms.

What does GPT-NeoX 20B on TPU v5p (per-chip, 95GB) cost per million tokens?

$2.44 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 GPT-NeoX 20B need on TPU v5p (per-chip, 95GB)?

20.0 GB for the weights at FP8 (E4M3), plus 70.9 GB of KV cache for 8 concurrent requests at 8K context — 90.9 GB total. Every extra concurrent user at long context costs more, so leave headroom.

Is TPU v5p (per-chip, 95GB) the right choice for GPT-NeoX 20B?

It works. Compare $2.44 per 1M tokens against the alternatives below — the cheapest option that fits the same workload is 8× RTX 4080 16GB at $0.0292, and the fastest is 4× Cerebras CS-3 (WSE-3) at 159120 tok/s.

Adjust the workload yourself

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

Open in calculator