GPUDatacenterReleased —

NVIDIA H20E

141 GB HBM3e

The H20E is a 2025 refresh of the H20, the Hopper-generation data center GPU NVIDIA built for the Chinese market under United States export limits on compute performance. It moves the GH100-based card from 96 GB of HBM3 to HBM3e, with 141 GB of memory reported, which suits inference on large language models where memory capacity rather than arithmetic throughput sets the limit. NVIDIA has not published a public specification sheet for it.

General

Manufacturer: NVIDIA
Model: H20E
Segment: Datacenter
GPU chip: GH100
Architecture: Hopper
Release date: —
Price: Search on Amazon

Memory

Memory size: 141 GB
Memory type: HBM3e
Bus width: —
Bandwidth: —
Memory clock: —

Clocks & silicon

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

Interface & APIs

Bus interface: —
SLI / multi-GPU: —
DirectX: N/A (compute-only)
OpenGL: N/A (compute-only)
Vulkan: N/A (compute-only)
Other APIs: CUDA (compute capability 9.0)

AI support

Precisions: FP8FP16BF16INT8INT4TF32FP64
AI accelerator: Fourth-generation Tensor Cores with Transformer Engine (FP8)
Peak AI TOPS: —
Notes: —