H100 CNX

Hopper NVIDIA H-Series Data Center Server HBM2e 2023 NVIDIA

Key Metrics

Launch Date: 2022-03-22
CUDA Cores: 14,592
Memory: 80 GB HBM2e
Memory Bandwidth: 2,040 GB/s
TDP/TBP: 350 W
Interface: PCIe 5.0 x16

Save this GPU, then open another GPU page to compare them side by side.

GPU Specs

Identity
Entry Type	Data Center GPU
Official Name	NVIDIA H100 CNX
GPU Codename	GH100
Architecture	Hopper
Launch Date	2022-03-22
Market Segment Notes	Data Center in NVIDIA H-Series
Core Specs
GPU Die	GH100
Process Node	TSMC 5 nm
Transistor Count	80 billion
Die Size	814 mm²
CUDA Cores	14592
SM Count	114
TMUs	456
ROPs	24
AI Units / Tensor Cores	456
Base Clock	690 MHz
Boost Clock	1,845 MHz
Pixel Fillrate	44.28 GP/s
Texture Fillrate	841.3 GT/s
FP32 Throughput	53.84 TFLOPS
FP16 Throughput	215.4 TFLOPS
FP64 Throughput	26.92 TFLOPS
Memory
Memory Size	80 GB
Memory Type	HBM2e
Memory Bus Width	5,120-bit
Memory Clock	1,593 MHz
Memory Bandwidth	2,040 GB/s
L2 Cache	50 MB
L1 Cache	250 KB
Board and Display
TDP / TBP	350 W
Recommended PSU	750 W
Power Connectors	8-pin EPS
Interface	PCIe 5.0 x16
Slot Width	Dual-slot, 267 mm length
Card Length	267 mm
Board Partner Model	Reference or OEM specification
Reference Design	NVIDIA reference design
Software
CUDA Compute Capability	9.0
OpenCL Support	OpenCL 3.0

Why It Mattered

H100 CNX belongs to the Hopper data-center generation, where transformer acceleration, HBM capacity, interconnect, and server form factor matter more than display features. It is tracked separately so accelerator buyers can distinguish PCIe, SXM, and OEM configurations.

Sources

Primary / manufacturer sources

NVIDIA Announces Hopper Architecture, the Next Generation of Accelerated ComputingOfficial March 22, 2022 launch release explicitly introducing the H100 CNX converged acceleratorOpen source

Independent reference sources

TechPowerUp GPU SpecsModel-level specification reference for this GPU entryOpen source

Reference databases

RightNow GPU DatabaseOpen GPU database used to cross-check specifications A reference database: good for cross-checking specifications, but not proof of the original release details on its own.Open source

Why It Mattered

H100 NVL 94 GB

H100 PCIe 80 GB

H100 PCIe 96 GB

H100 SXM5 64 GB

Primary / manufacturer sources

Independent reference sources

Reference databases