L40S vs RTX 4090: Which GPU Should You Choose?

This page settles whether L40S or RTX 4090 is the better rental for your workload, using observed cloud pricing, published hardware specifications, and workload-level cost estimates.

Updated 2026-08-23 · refreshed daily

Choose L40S when its larger memory lets the job stay on one GPU; choose RTX 4090 when the job fits one card and rental price matters, given an A-to-B on-demand price ratio of 2.86, except when L40S avoids a multi-GPU split.

Run one short A/B benchmark of your representative job at the L40S on-demand minimum of $0.80 and the RTX 4090 on-demand minimum of $0.28, verifying both rates in the live table first.

How much GPU memory must your workload fit within? Choose RTX 4090 for FP8, BF16, or FP32 workloads within 24GB VRAM; move to L40S when the job exceeds that memory ceiling but still fits FP8, BF16, or FP32 workloads within 48GB VRAM, otherwise plan a multi-GPU configuration.
Is finishing sooner worth the difference in per-job cost? Choose L40S when the deadline is firm, since its effective cost-per-job ratio versus RTX 4090 is 1.3; choose RTX 4090 when turnaround is flexible and minimizing hourly spend matters more.
Can you reliably obtain enough matching GPUs when needed? Book a short test rental for the full required GPU count from a currently available listing before committing the full run.

Today's prices

USD/hr · 1× GPU · lowest on-demand price per provider

ProviderL40SRTX 4090
Vast.ai $0.80 $0.28
RunPod $0.99 $0.74
DataCrunch $1.37 Not offered
Crusoe $1.50 · Not reported Not offered

Prices refresh daily. Marlin always books the cheapest available. Try Marlin →
2.86×
The on-demand premium makes L40S worthwhile when its larger memory avoids extra GPUs.
1.30×
At on-demand rates, L40S delivers a lower effective cost per job than RTX 4090.
$584/mo
Monthly baseline for running one L40S continuously at the on-demand minimum.

What your workload needs

Find your parameter count on the model's Hugging Face card. Serving maps to the inference rows, training to the fine-tuning rows.

WorkloadVRAM neededL40SRTX 4090Cheapest today
70B Q4 inference 42 GB 1× $0.80/hr 2× $0.28/hr RTX 4090 ×2 → $0.56/hr (est.)
70B FP16 inference 168 GB 4× $0.80/hr 7× $0.28/hr RTX 4090 ×7 → $1.96/hr (est.)
70B QLoRA 46 GB 1× $0.80/hr 2× $0.28/hr RTX 4090 ×2 → $0.56/hr (est.)
7B full fine-tune 112 GB 3× $0.80/hr 5× $0.28/hr RTX 4090 ×5 → $1.4/hr (est.)
70B FP8 training 154 GB 4× $0.80/hr 7× $0.28/hr RTX 4090 ×7 → $1.96/hr (est.)
Downshift recommendation 42 GB 1× $0.80/hr 2× $0.28/hr L40S ×1 → $0.8/hr

What the numbers say

L40S carries the higher on-demand minimum, with an A-to-B on-demand price ratio of 2.86. That makes RTX 4090 the default when the workload fits one card and hourly spend is the priority, while L40S earns its premium when additional memory avoids a multi-GPU configuration.

L40S’s published BF16 peak implies an A-to-B performance ratio of 2.19, but it is not measured workload throughput. Applied to the on-demand pricing comparison, that produces an effective cost-per-job ratio of 1.3 versus RTX 4090, suggesting L40S can complete equivalent compute for less when real-world utilization and software efficiency track the datasheet figures.

For RTX 4090, the baseline is FP8, BF16, or FP32 workloads within 24GB VRAM; the exception applies when model weights, KV cache, activations, or training state exceed that single-card ceiling. L40S supports the same precision choices with a larger VRAM ceiling, so choose it when that headroom keeps the workload on one GPU; otherwise RTX 4090 is the rational choice.

Before you rent

Data egress charges sit outside the headline GPU rate; verify the provider's current per-GB outbound transfer fee on its live pricing page before estimating total cost.
A workload that exceeds single-GPU VRAM forces multi-GPU partitioning, increasing the rented GPU count and potentially adding communication overhead; budget against the full configuration, not one card.
Interruptible capacity can disappear mid-run, adding checkpoint recovery time and duplicate compute; use it only for workloads that can resume cleanly.
Billing granularity can make short experiments cost more than runtime alone suggests because minimum charges may round usage up; each provider's pricing page lists the applicable minimum charge.

Marlin matches your workload requirements to the lowest-priced suitable GPU across major cloud providers.

Coverage is limited to the listed providers and published hardware specifications, while live availability and prices can change without notice. Workload costs and performance comparisons are modeled from VRAM formulas, on-demand rates, and datasheet peaks rather than measured end-to-end benchmarks. https://vast.ai/pricing · https://crusoe.ai/cloud/pricing · https://www.runpod.io/pricing · https://datacrunch.io/pricing · https://resources.nvidia.com/en-us-l40s/l40s-datasheet-28413 · https://www.nvidia.com/en-us/geforce/graphics-cards/40-series/rtx-4090/

Stop comparing. Start running.

Marlin matches your workload to the lowest-priced suitable GPU across supported providers.

Get started with Marlin