NVIDIA48 GB VRAM864 GB/s$0.7/hr

DeepSeek Coder V2 236B (MoE) on L40S 48GB: Does It Fit, How Fast, What It Costs

DeepSeek Coder V2 236B (MoE) at AWQ 4-bit needs 129.8 GB for its weights and 142.4 GB in total at 8K context with 8 concurrent users. That means it needs 4× L40S 48GB. Every figure is computed by tokcalc's open-source capacity model.

Fits?
Yes, 4×
142.4 GB needed / 192 GB available
Decode speed
145.5
tok/s per stream
Aggregate
6983
tok/s @ 8 concurrent
Cost / 1M tokens
$0.1114
4× L40S 48GB

L40S 48GB has no NVLink, so 4-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 GB142.4 GB0.0000.000——✗
2× L40S96 GB142.4 GB0.0000.000——✗
4× L40S192 GB142.4 GB145.569831.91 s$0.1114✓
8× L40S384 GB142.4 GB291.027932972.5 ms$0.0557✓
Concurrency vs context length
On 4× L40S 48GB at AWQ 4-bit. KV cache scales linearly with context, so this is the number of users your GPUs can actually serve.
ContextConcurrent usersKV per requestFits?
4K78807.5 MB✓ fits
8K391.58 GB✓ fits
33K96.31 GB✓ fits
66K412.62 GB✓ fits
131K225.23 GB✓ fits
Other GPUs that run DeepSeek Coder V2 236B (MoE)
Same workload, same quantization — ranked by value (tokens per dollar-hour).
GPUGPUsAggregate tok/sCost / 1M tok
RTX 5090 32GB8×57934$0.0192
A100 80GB SXM42×4041$0.1650
H100 SXM5 80GB2×6769$0.2052
AMD MI300X 192GB1×3150$0.2205
H200 SXM5 141GB2×9699$0.2291
AMD MI325X 288GB1×3566$0.2727
Frequently asked questions

Can DeepSeek Coder V2 236B (MoE) run on L40S 48GB?

Yes, but it needs 4× L40S 48GB. At AWQ 4-bit the total footprint is 142.4 GB (129.8 GB of weights plus 12.6 GB of KV cache at 8K × 8 concurrent users), against 192 GB of available VRAM across 4 GPUs.

How fast is DeepSeek Coder V2 236B (MoE) on L40S 48GB?

145.5 tokens/sec per stream and 6983 tok/s aggregate with continuous batching at 8 concurrent users. Time to first token is 1.91 s, and inter-token latency is 6.9 ms.

What does DeepSeek Coder V2 236B (MoE) on L40S 48GB cost per million tokens?

$0.1114 per 1M output tokens at 4× 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 DeepSeek Coder V2 236B (MoE) need on L40S 48GB?

129.8 GB for the weights at AWQ 4-bit, plus 12.6 GB of KV cache for 8 concurrent requests at 8K context — 142.4 GB total. Every extra concurrent user at long context costs more, so leave headroom.

Is L40S 48GB the right choice for DeepSeek Coder V2 236B (MoE)?

It works. Compare $0.1114 per 1M tokens against the alternatives below — the cheapest option that fits the same workload is 8× RTX 5090 32GB at $0.0192, and the fastest is 4× Cerebras CS-3 (WSE-3) at 161646 tok/s.

Adjust the workload yourself

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

Open in calculator