GPUDatacenterReleased —

NVIDIA H20

96 GB HBM3 · 6144-bit

The H20 is a Hopper-based data-center GPU that NVIDIA created for China to comply with US export controls, reaching customers in volume in 2024. It keeps 96 GB of HBM3 memory and full NVLink bandwidth but has compute throughput cut far below the H100, making it better suited to AI inference than training. After a US licensing requirement in April 2025 and Chinese discouragement of purchases, NVIDIA halted H20 production in August 2025, and the H200 was later approved for export to China in its place.

General

Manufacturer: NVIDIA
Model: H20
Segment: Datacenter
GPU chip: GH100
Architecture: Hopper
Release date: —
Price: Search on Amazon

Memory

Memory size: 96 GB
Memory type: HBM3
Bus width: 6144-bit
Bandwidth: 4,000 GB/s
Memory clock: —

Clocks & silicon

GPU clock: —
Boost clock: —
Shader units: —
Tensor / AI cores: —
RT cores: —
Transistors: 80 billion
Technology (nm): 4 nm
TSMC 4N
TDP / TBP: 400 W

Interface & APIs

Bus interface: SXM5 (PCIe 5.0 x16 host)
SLI / multi-GPU: NVLink 4.0 (900 GB/s)
DirectX: N/A (compute only)
OpenGL: N/A (compute only)
Vulkan: N/A
Other APIs: CUDA (compute capability 9.0), OpenCL 3.0

AI support

Precisions: FP8FP16BF16INT8INT4TF32FP64
AI accelerator: 4th Gen Tensor Cores + Transformer Engine
Peak AI TOPS: 296 TOPS
Notes: Dense throughput of about 296 TFLOPS FP8 and 296 TOPS INT8, 148 TFLOPS FP16 and BF16, 74 TFLOPS TF32, 44 TFLOPS FP32 and 1 TFLOPS FP64. Hopper has no INT4 support. Supports Multi-Instance GPU (MIG) with up to 7 instances.