Report an Error

NVIDIA GB102

NVIDIA's GB102 GPU uses the Blackwell architecture and is made using a 5 nm production process at TSMC. GB102 does not support DirectX. For GPU compute applications, OpenCL version 3.0 and CUDA 10.1 can be used. It features 18432 shading units, 576 texture mapping units and 24 ROPs. Also included are 576 tensor cores which help improve the speed of machine learning applications.
Further reading: Blackwell Architecture Whitepaper

Graphics Processor

Released
Nov 2024
GPU Name
GB102
Codename
NV190
Architecture
Blackwell
Foundry
TSMC
Process Type
4NP FinFET
Process Size
5 nm
Transistors
104,000 million
Die Size
unknown

Graphics Features

DirectX
N/A
OpenGL
N/A
OpenCL
3.0
Vulkan
N/A
CUDA
10.1
Shader Model
N/A
Tensor Cores
5th Gen
PureVideo HD
VP11
VDPAU
Feature Set K

Render Config

Shading Units
18432
TMUs
576
ROPs
24
SM Count
132
FP16 Units
73,728
FP64 Units
9216
INT32 Units
9216
Tensor Cores
576
Tex L1 Cache
64 KB per SM
L1 Cache
256 KB per SM
L2 Cache
61440 KB
Max. TDP
1000 W
SFUs
2304
TPCs
72
GPCs
8

All Blackwell GPUs

NVIDIA GPU Architecture History

Graphics cards using the NVIDIA GB102 GPU

Name Chip Memory Shaders TMUs ROPs Base Clock Boost Clock Memory Clock
NVIDIA B100
96 GB 16896 528 24 1665 MHz 1837 MHz 2000 MHz

GB102 GPU Notes

Tensor Cores: 5th Gen
NVENC: No Support
NVDEC: No Support
PureVideo HD: VP11
VDPAU: Feature Set K

Most specs unknown
Placeholder
May 20th, 2025 16:58 CDT change timezone

Latest GPU Drivers

New Forum Posts

Popular Reviews

Controversial News Posts