GeekRadar / The comparison desk

COMPARE.

GPU

NVIDIA

NVIDIA H100 NVL 94 GB

Graphics processor

Video memory: 94 GB^*
Memory bus: 6,016 bit^*
Rated memory bandwidth: 3,950 GB/s^*

No matched benchmark results are available for this entry yet.

Processing & platform

Architecture: Hopper^*
Process node: 5 nm^*
GPU cores: 16,896^*
Graphics interface: PCIe 5.0 x16^*
Reported chip die area: 814 mm²^*
Transistors: 80,000 million^*
GPU L2 cache: 60 MB^*
Graphics processor: GH100^*
Texture mapping units: 528^*
Raster output units: 24^*
Streaming multiprocessors: 132^*
Tensor cores: 528^*
Chip foundry: TSMC^*
Card or module format: Dual-slot^*
Declared GPU TDP: 400 W^*
Suggested system PSU: 800 W^*
External power connectors: 8-pin EPS^*
OpenCL: 3.0^*
CUDA compute capability: 9.0^*

Memory & storage

Memory options: 94^*
Video memory: 94 GB^*
Video memory type: HBM3^*
Memory bus: 6,016 bit^*
Rated memory bandwidth: 3,950 GB/s^*

Power specifications

Thermal design power: 400 W^*

System configuration

Detailed specifications

Specifications 1 · 27 fields

GPU architecture — NVIDIA H100 NVL — 94 GB HBM3

Hopper^*

Calculated theoretical peak: BF16 tensor dense per GPU — NVIDIA H100 NVL — 94 GB HBM3

835.5^* TFLOPS

Scope and conditions

Calculation: 1671 / 2^*

Theoretical peak: BF16 tensor with sparsity per GPU — NVIDIA H100 NVL — 94 GB HBM3

1,671^* TFLOPS

Calculated theoretical peak: FP16 tensor dense per GPU — NVIDIA H100 NVL — 94 GB HBM3

835.5^* TFLOPS

Scope and conditions

Calculation: 1671 / 2^*

Theoretical peak: FP16 tensor with sparsity per GPU — NVIDIA H100 NVL — 94 GB HBM3

1,671^* TFLOPS

Theoretical peak: FP32 non-tensor dense per GPU — NVIDIA H100 NVL — 94 GB HBM3

60^* TFLOPS

Native hardware support: FP4 tensor — NVIDIA H100 NVL — 94 GB HBM3

No^*

Theoretical peak: FP64 non-tensor dense per GPU — NVIDIA H100 NVL — 94 GB HBM3

30^* TFLOPS

Theoretical peak: FP64 tensor dense per GPU — NVIDIA H100 NVL — 94 GB HBM3

60^* TFLOPS

Native hardware support: FP6 tensor — NVIDIA H100 NVL — 94 GB HBM3

No^*

Calculated theoretical peak: FP8 tensor dense per GPU — NVIDIA H100 NVL — 94 GB HBM3

1,670.5^* TFLOPS

Scope and conditions

Calculation: 3341 / 2^*

Theoretical peak: FP8 tensor with sparsity per GPU — NVIDIA H100 NVL — 94 GB HBM3

3,341^* TFLOPS

Calculated theoretical peak: INT8 tensor dense per GPU — NVIDIA H100 NVL — 94 GB HBM3

1,670.5^* TOPS

Scope and conditions

Calculation: 3341 / 2^*

Theoretical peak: INT8 tensor with sparsity per GPU — NVIDIA H100 NVL — 94 GB HBM3

3,341^* TOPS

Calculated theoretical peak: TF32 tensor dense per GPU — NVIDIA H100 NVL — 94 GB HBM3

417.5^* TFLOPS

Scope and conditions

Calculation: 835 / 2^*

Theoretical peak: TF32 tensor with sparsity per GPU — NVIDIA H100 NVL — 94 GB HBM3

835^* TFLOPS

Physical configuration — NVIDIA H100 NVL — 94 GB HBM3

NVL (PCIe dual-slot, NVLink bridge)^*

GPU interconnect bandwidth (NVLink 4 (three NVLink bridges, two cards)) bidirectional — NVIDIA H100 NVL — 94 GB HBM3

600^* GB/s

GPU interconnect bandwidth per direction — NVIDIA H100 NVL — 94 GB HBM3

300^* GB/s

Scope and conditions

Calculation: 600 / 2^*

PCIe interface declaration — NVIDIA H100 NVL — 94 GB HBM3

PCIe Gen5: 128GB/s^*

Maximum board power — NVIDIA H100 NVL — 94 GB HBM3

400^* W

Peak memory bandwidth per GPU — NVIDIA H100 NVL — 94 GB HBM3

3,938^* GB/s

Memory interface width per GPU — NVIDIA H100 NVL — 94 GB HBM3

6,016^* bit

Memory capacity per GPU — NVIDIA H100 NVL — 94 GB HBM3

94^* GB

Memory reported in a documented driver sample — NVIDIA H100 NVL — 94 GB HBM3

95,830^* MiB

Scope and conditions

Configuration: Documented driver sample; not nominal VRAM capacity.^*

Memory technology — NVIDIA H100 NVL — 94 GB HBM3

HBM3^*

GPU model — NVIDIA H100 NVL — 94 GB HBM3

NVIDIA H100 NVL^*

Specifications 2 · 30 fields

proc transistors: 80,000 million^*
proc die size: 814^* mm²
cores: 16,896^*
tensor cores: 528^*
sm count: 132^*
base clock: 1,080^*
boost clock: 1,785^*
memory clock: 1,310^*
ram: 94^*
name: NVIDIA H100 NVL 94 GB^*
ram type: HBM3^*
bus width: 6,016^*
bus interface: PCIe 5.0 x16^*
ram bandwidth: 3.94 TB/s^*
fp16 performance: 241.3^*
fp32 performance: 60.32^*
fp64 performance: 30.16^*
tdp: 400^* W
suggested psu: 800^* W
slot width: Dual-slot^*
power connectors: 8-pin EPS^*
api cuda: 9^*
is retail board: No^*
gpi value: 57.23^*
vendor: nvd^*
manufacturer: NVIDIA^*
arch name: Hopper^*
card release date: 2023-03-21^*
proc foundry: TSMC^*
proc process size: 5^* nm

Specifications 3 · 35 fields

CUDA compute capability

9.0^*

OpenCL

3.0^*

Reported theoretical BF16 peak

965.1^* TFLOPS

Scope and conditions

Arithmetic density: Not stated^*

Graphics interface

PCIe 5.0 x16^*

Memory bus

6,016^* bit

Reported shader processors

16,896^*

Reported theoretical FP16 peak

120.6^* TFLOPS

Scope and conditions

Arithmetic density: Not stated^*

Reported theoretical FP32 peak

60.32^* TFLOPS

Scope and conditions

Arithmetic density: Not stated^*

Reported theoretical FP64 peak

30.16^* TFLOPS

Scope and conditions

Arithmetic density: Not stated^*

Declared L1 cache configuration

256 KB (per SM)^*

Declared GPU L2 cache

60 MB^*

Reported display outputs

No outputs^*

Theoretical pixel fill rate

42.84 GPixel/s^*

External power connectors

8-pin EPS^*

Reported chip die area

814 mm²^*

Chip foundry

TSMC^*

Process node declaration

5 nm^*

Transistors

80.0 billion^*

Graphics processor

GH100^*

Reported video memory

94^* GB

Rated memory bandwidth

3.95 TB/s^*

Video memory type

HBM3^*

Raster output units

24^*

Reported shading units

16,896^*

Card or module format

Dual-slot^*

Streaming multiprocessors

132^*

Suggested system PSU

800 W^*

Declared GPU TDP

400 W^*

Tensor cores

528^*

Theoretical texture fill rate

942.5 GTexel/s^*

Reported theoretical TF32 peak

482.5^* TFLOPS

Scope and conditions

Arithmetic density: Not stated^*

Texture mapping units

528^*

GPU cores

16,896^*

Video memory

94^* GB

Video memory type

HBM3^*

* Provisional data