H100 SXM vs L40S: Which GPU Should You Choose?

This page settles whether H100 SXM or L40S better fits your workload, using provider pricing, vendor specifications, and transparent workload estimates.

Updated 2026-08-23 · refreshed daily

H100 SXM wins when its larger VRAM keeps a workload on one GPU, while L40S wins on hourly cost for workloads that fit, given the H100 SXM-to-L40S on-demand price ratio of 2.17:1, except when NVLink is required.

Use the live table to verify that the H100 SXM and L40S on-demand minimums are still $1.74/hr and $0.80/hr before committing.

How much GPU memory does your workload actually require? Choose L40S for FP8, BF16, or FP32 workloads within 48GB VRAM; choose H100 SXM for FP8, BF16, or FP32 workloads within 80GB VRAM when the workload exceeds L40S's VRAM ceiling, and use a multi-GPU configuration if it exceeds both.

Is finishing sooner worth the difference in total job cost? For a hard deadline, choose H100 SXM only if a representative benchmark confirms it finishes sooner; otherwise choose L40S, since H100 SXM has a modeled effective cost-per-job ratio of 0.8 relative to L40S.

Do you need guaranteed GPU capacity at a specific time? Request a capacity reservation for your required GPU and target window before choosing hardware.

Today's prices

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

Provider	H100 SXM	L40S
Vast.ai	$1.74	$0.80
Hyperstack	$3.20 · Not reported	Not offered
DataCrunch	$3.25	$1.37
RunPod	$3.29	$0.99
Crusoe	$3.90 · Not reported	$1.50 · Not reported
Together AI	$3.99 · Not reported	Not offered
Paperspace	$5.95 · Not reported	Not offered

see all H100 prices →

Prices refresh daily. Marlin always books the cheapest available. Try Marlin →

2.17×

The on-demand price premium H100 SXM must justify through workload fit

0.80×

L40S retains the lower modeled cost per job at on-demand rates

$1270/mo

Baseline monthly budget for running one H100 SXM continuously at its 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.

Workload	VRAM needed	H100 SXM	L40S	Cheapest today
70B Q4 inference	42 GB	1× $1.74/hr	1× $0.80/hr	L40S ×1 → $0.8/hr
70B FP16 inference	168 GB	3× $1.74/hr	4× $0.80/hr	L40S ×4 → $3.2/hr (est.)
70B QLoRA	46 GB	1× $1.74/hr	1× $0.80/hr	L40S ×1 → $0.8/hr
7B full fine-tune	112 GB	2× $1.74/hr	3× $0.80/hr	L40S ×3 → $2.4/hr (est.)
70B FP8 training	154 GB	2× $1.74/hr	4× $0.80/hr	L40S ×4 → $3.2/hr (est.)
Downshift recommendation	42 GB	1× $1.74/hr	1× $0.80/hr	L40S ×1 → $0.8/hr

What the numbers say

H100 SXM carries an on-demand price ratio of 2.17:1 versus L40S, making L40S the cost-first choice when the workload fits its VRAM. Choose H100 SXM when its larger memory or NVLink support avoids an unsuitable configuration.

H100 SXM has a vendor-rated peak BF16 performance ratio of 2.73 versus L40S, not a measured workload result. Using that peak with on-demand pricing yields a modeled effective cost-per-job ratio of 0.8 for H100 SXM relative to L40S, so verify actual job economics with a representative benchmark.

L40S covers FP8, BF16, or FP32 workloads within 48GB VRAM, while H100 SXM adds a larger single-GPU memory ceiling for the same precision modes. If model weights and runtime memory exceed L40S’s ceiling but still fit on one H100 SXM, choose H100 SXM; otherwise, L40S is the rational choice.

Before you rent

Data egress can raise the total bill beyond GPU hours; verify the provider's network pricing for your region and expected transfer volume.

Exceeding single-GPU VRAM can force a multi-GPU configuration, increasing GPU count and communication overhead.

Spot rows marked interruptible expose jobs to restarts, which can erase rate savings when the workload cannot resume from a checkpoint.

Billing granularity can make a short experiment cost more than its runtime suggests; find the minimum charge on each provider's pricing page when estimating the billed amount.

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

Data comes from provider pricing pages and vendor datasheets, with source coverage and freshness shown separately; prices and availability can change without notice. Performance and workload costs are modeled from published specifications, on-demand rates, and memory formulas rather than measured end-to-end benchmarks. https://www.tensordock.com/host-pricing · https://www.coreweave.com/pricing · https://vast.ai/pricing · https://datacrunch.io/pricing · https://resources.nvidia.com/en-us-tensor-core/nvidia-tensor-core-gpu-datasheet · https://resources.nvidia.com/en-us-l40s/l40s-datasheet-28413 · https://www.runpod.io/pricing · https://lambda.ai/service/gpu-cloud · https://www.hyperstack.cloud/gpu-pricing · https://www.together.ai/pricing · https://crusoe.ai/cloud/pricing

Stop comparing. Start running.

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

Get started with Marlin