Skip to content
Cloud Provider Benchmarks

Datasets

Data-centre GPU specifications

Memory, memory bandwidth, dense FP8/BF16/FP32 throughput, interconnect and power for NVIDIA H100, H200 and L40S variants, from vendor datasheets.

Version
1.0
Updated
Rows
6
License
CC BY 4.0

JSONCSV

Vendor datasheet values; dense (non-sparse) throughput.

Fields

FieldDescription
gpuGPU id
variantVariant id
nameVariant name
architectureGPU architecture
form_factorForm factor
memory_gbDevice memory, GB
memory_typeMemory technology
bandwidth_gbsMemory bandwidth, GB/s
fp8_dense_tflopsDense FP8 Tensor Core TFLOPS
bf16_dense_tflopsDense BF16 Tensor Core TFLOPS
fp32_tflopsFP32 TFLOPS
nvlink_gbsNVLink bandwidth, GB/s (null if none)
tdpPower envelope
sourcePrimary datasheet URL

Preview

First 6 of 6 rows.

gpuvariantnamearchitectureform_factormemory_gbmemory_typebandwidth_gbsfp8_dense_tflopsbf16_dense_tflopsfp32_tflopsnvlink_gbstdpsource
nvidia-h100h100-sxmH100 SXM5HopperSXM580HBM33350197998967900Up to 700 Whttps://www.nvidia.com/en-us/data-center/h100/
nvidia-h100h100-pcieH100 PCIeHopperPCIe dual-slot80HBM2e2000151375651600300–350 Whttps://www.nvidia.com/en-us/data-center/h100/
nvidia-h100h100-nvlH100 NVLHopperPCIe dual-slot94HBM33900167083560600350–400 Whttps://www.nvidia.com/en-us/data-center/h100/
nvidia-h200h200-sxmH200 SXMHopperSXM5141HBM3e4800197998967900Up to 700 Whttps://www.nvidia.com/en-us/data-center/h200/
nvidia-h200h200-nvlH200 NVLHopperPCIe dual-slot141HBM3e4800167083560900Up to 600 Whttps://www.nvidia.com/en-us/data-center/h200/
nvidia-l40sl40sL40SAda LovelacePCIe dual-slot48GDDR6 (ECC)86473336291.6—350 Whttps://www.nvidia.com/en-us/data-center/l40s/

Changelog

  • v1.0 · — Initial release.