NVIDIA48 GB VRAM864 GB/s$0.7/hr

Qwen 3 32B on L40S 48GB: Does It Fit, How Fast, What It Costs

Qwen 3 32B at FP8 (E4M3) needs 32.0 GB for its weights and 49.2 GB in total at 8K context with 8 concurrent users. That means it needs 2× L40S 48GB. Every figure is computed by tokcalc's open-source capacity model.

Fits?
Yes, 2×
49.2 GB needed / 96 GB available
Decode speed
44.75
tok/s per stream
Aggregate
1074
tok/s @ 8 concurrent
Cost / 1M tokens
$0.3621
2× L40S 48GB

L40S 48GB has no NVLink, so 2-way tensor parallelism runs over PCIe. tokcalc applies a flat 0.85× multi-GPU efficiency factor for this, which is realistic for NVLink-equipped datacenter cards but optimistic for PCIe scale-out — treat the throughput and cost figures for this configuration as an upper bound, and measure before committing.

What changes as you add L40S 48GBs
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× L40S48 GB49.2 GB0.0000.000——✗
2× L40S96 GB49.2 GB44.7510745.78 s$0.3621✓
4× L40S192 GB49.2 GB89.5042962.90 s$0.1810✓
8× L40S384 GB49.2 GB179.0171851.47 s$0.0905✓
Concurrency vs context length
On 2× L40S 48GB 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?
4K591.07 GB✓ fits
8K292.15 GB✓ fits
33K78.59 GB✓ fits
66K317.18 GB✓ fits
131K134.36 GB✓ fits
Other GPUs that run Qwen 3 32B
Same workload, same quantization — ranked by value (tokens per dollar-hour).
GPUGPUsAggregate tok/sCost / 1M tok
RTX 5090 32GB2×2228$0.1247
AMD MI300X 192GB1×1938$0.3584
AMD MI325X 288GB1×2194$0.4432
A100 80GB SXM41×731.3$0.4558
Intel Gaudi 3 128GB1×1207$0.4604
H100 SXM5 80GB1×1225$0.5670
Frequently asked questions

Can Qwen 3 32B run on L40S 48GB?

Yes, but it needs 2× L40S 48GB. At FP8 (E4M3) the total footprint is 49.2 GB (32.0 GB of weights plus 17.2 GB of KV cache at 8K × 8 concurrent users), against 96 GB of available VRAM across 2 GPUs.

How fast is Qwen 3 32B on L40S 48GB?

44.75 tokens/sec per stream and 1074 tok/s aggregate with continuous batching at 8 concurrent users. Time to first token is 5.78 s, and inter-token latency is 22.3 ms.

What does Qwen 3 32B on L40S 48GB cost per million tokens?

$0.3621 per 1M output tokens at 2× L40S 48GB ($0.7000/hr). That is the marginal hardware cost — add idle capacity, ops, and egress for the real figure.

How much VRAM does Qwen 3 32B need on L40S 48GB?

32.0 GB for the weights at FP8 (E4M3), plus 17.2 GB of KV cache for 8 concurrent requests at 8K context — 49.2 GB total. Every extra concurrent user at long context costs more, so leave headroom.

Is L40S 48GB the right choice for Qwen 3 32B?

It works. Compare $0.3621 per 1M tokens against the alternatives below — the cheapest option that fits the same workload is 8× RTX 3080 10GB at $0.0294, and the fastest is 2× Cerebras CS-3 (WSE-3) at 24863 tok/s.

Adjust the workload yourself

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

Open in calculator