Skip to content

H100 SXM vs H100 PCIe

Side-by-side comparison of the NVIDIA H100 SXM and the NVIDIA H100 PCIe for AI inference workloads.

Specifications

SpecH100 SXMH100 PCIe
Generationhopperhopper
Memory TypeHBM3HBM3
VRAM80 GB80 GB
Memory Bandwidth3,350 GB/s2,000 GB/s
BF16 TFLOPS990756
FP16 TFLOPS990756
FP8 TFLOPS1,9791,513
INT8 TOPS1,9791,513
TDP700 W350 W
Interconnectnvlinkpcie
NVLink Bandwidth900 GB/sN/A
Max GPUs per Node88
PCIe GenGen 5Gen 5
CUDA Compute Capability99

Pricing

H100 SXM

ProviderOn-DemandReservedSpot
runpod$4.18/hr-$3.29/hr
lambda$2.49/hr$1.89/hr-
coreweave$3.79/hr$2.57/hr-
aws$5.12/hr$3.59/hr-
gcp$4.85/hr$3.40/hr-
azure$4.98/hr$3.49/hr-
vast ai$3.40/hr-$2.50/hr
tensordock$3.29/hr-$2.49/hr
fluidstack$2.85/hr-$2.10/hr

H100 PCIe

ProviderOn-DemandReservedSpot
runpod$3.09/hr-$2.39/hr
lambda$2.29/hr--
tensordock$2.59/hr-$1.89/hr
vast ai$2.80/hr-$2.10/hr

Cheapest available rate: H100 SXM at $1.89/hr vs H100 PCIe at $1.89/hrSame price

Efficiency Metrics

TFLOPS / Watt

1.4

H100 SXM

vs

2.2

H100 PCIe

BF16

VRAM / Dollar

42.3

H100 SXM

vs

42.3

H100 PCIe

GB/$/hr

Bandwidth / Watt

4.8

H100 SXM

vs

5.7

H100 PCIe

GB/s/W

Models (FP16, 1 GPU)

182.0

H100 SXM

vs

182.0

H100 PCIe

Model Compatibility (FP16, Single GPU)

Only on H100 SXM (0)

None

Both (182)

  • Yi 1.5 34B
  • Yi 1.5 9B
  • Yi Coder 9B
  • GTE Qwen2 7B
  • Marco O1
  • Qwen 1.5 MoE A2.7B
  • Qwen 2 Audio 7B
  • Qwen 2.5 14B
  • Qwen 2.5 32B
  • Qwen 2.5 3B
  • Qwen 2.5 Coder 32B
  • OLMo 2 13B
  • OLMo 2 7B
  • Amazon Nova Lite
  • OpenELM 3B
  • BGE Large EN v1.5
  • BGE M3
  • Baichuan 2 13B
  • OctoCoder 15B
  • StarCoder2 15B
  • +162 more

Only on H100 PCIe (0)

None

Summary

The H100 SXM (hopper generation) offers 80GB of HBM3 with 990 BF16 TFLOPS and 3,350 GB/s memory bandwidth at 700W TDP.

The H100 PCIe (hopper generation) offers 80GB of HBM3 with 756 BF16 TFLOPS and 2,000 GB/s memory bandwidth at 350W TDP.

More GPU Comparisons