H100 PCIe vs H100 SXM: Which GPU Should You Choose?

This page settles whether H100 PCIe or H100 SXM better fits your workload, based on observed cloud pricing and published NVIDIA specifications.

Updated 2026-08-23 · refreshed daily

At an on-demand price ratio of 1.23, H100 PCIe wins when hourly spend matters most, while H100 SXM wins for performance-bound workloads, unless its extra throughput cannot shorten the job.

Run one short representative-job A/B benchmark at the on-demand minimums of $2.14 for H100 PCIe and $1.74 for H100 SXM, verifying both rates in the live table before launch.

Does your workload fit on a single GPU? Both qualify: H100 PCIe supports FP8, BF16, or FP32 workloads within 80GB VRAM, and H100 SXM supports FP8, BF16, or FP32 workloads within 80GB VRAM; if your workload exceeds either threshold, move to a multi-GPU configuration.
Is finishing sooner worth the difference in cost per job? Choose H100 SXM when deadline matters; the 1.61 effective cost-per-job ratio for H100 PCIe versus H100 SXM indicates that it also has the lower modeled job cost.
Can your workload tolerate interruptions? Run one short, checkpointed H100 SXM test rental on an interruptible instance before committing.

Today's prices

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

ProviderH100 PCIeH100 SXM
Vast.ai $2.14 $1.74
Hyperstack $2.50 · Not reported $3.20 · Not reported
RunPod $2.89 $3.29
DataCrunch Not offered $3.25
Crusoe Not offered $3.90 · Not reported
Together AI Not offered $3.99 · Not reported
Paperspace Not offered $5.95 · Not reported

see all H100 prices → · see all H100 prices →

Prices refresh daily. Marlin always books the cheapest available. Try Marlin →
1.23×
On-demand price edge favors H100 PCIe
1.61×
H100 SXM delivers the lower effective cost per job
$1562/mo
Monthly cost to keep H100 PCIe running 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.

WorkloadVRAM neededH100 PCIeH100 SXMCheapest today
70B Q4 inference 42 GB 1× $2.14/hr 1× $1.74/hr H100 SXM ×1 → $1.74/hr
70B FP16 inference 168 GB 3× $2.14/hr 3× $1.74/hr H100 SXM ×3 → $5.22/hr (est.)
70B QLoRA 46 GB 1× $2.14/hr 1× $1.74/hr H100 SXM ×1 → $1.74/hr
7B full fine-tune 112 GB 2× $2.14/hr 2× $1.74/hr H100 SXM ×2 → $3.48/hr (est.)
70B FP8 training 154 GB 2× $2.14/hr 2× $1.74/hr H100 SXM ×2 → $3.48/hr (est.)
Downshift recommendation 42 GB 1× $2.14/hr 1× $1.74/hr H100 SXM ×1 → $1.74/hr

What the numbers say

H100 PCIe has an on-demand price ratio of 1.23 relative to H100 SXM, making it the stronger choice when comparable jobs run for similar durations; when faster completion changes billed runtime, hourly price alone should not decide.

H100 SXM is the performance-led choice: vendor peak specifications imply a 0.76 performance ratio for H100 PCIe versus H100 SXM, which yields an on-demand effective cost-per-job ratio of 1.61. These are modeled results from datasheet peaks, not measured workload throughput, so validate them with your own job.

H100 PCIe and H100 SXM share the same workload threshold: FP8, BF16, or FP32 workloads within 80GB VRAM. The exception is a job that must span multiple GPUs or whose runtime depends heavily on inter-GPU communication; choose H100 SXM for that case, and choose H100 PCIe when the workload fits one GPU and hourly spend is the priority.

Before you rent

Data egress fees can raise the total bill; verify each provider's current outbound-transfer rate and included allowance before moving checkpoints or results.
Multi-GPU scaling efficiency can change total cost because synchronization overhead may add GPU-hours when a workload does not parallelize cleanly.
Interruptible instances can disappear mid-job, turning weak checkpointing into lost compute time and a longer completion window.
Billing granularity can make short experiments cost more than their runtime suggests; use each provider's pricing page to find the minimum charge applied to a brief rental.

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

Coverage is limited to the listed provider offers, and pricing and availability can change after publication. Performance, VRAM needs, GPU counts, and cost per job are modeled from published specifications and workload assumptions rather than measured end-to-end benchmarks. https://vast.ai/pricing · https://lambda.ai/service/gpu-cloud · https://www.coreweave.com/pricing · https://datacrunch.io/pricing · https://resources.nvidia.com/en-us-tensor-core/nvidia-tensor-core-gpu-datasheet · https://www.runpod.io/pricing · 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