NVIDIA H20 141GB HBM3e
Overview
The NVIDIA H20-3e (also called "H20E") is the HBM3e memory refresh of the China-export H20, announced January 2025. Its compute is identical to the original H20 96GB (export-capped Hopper, 4nm): dense peaks of 44 TFLOPS FP32, 1 TFLOPS FP64, 74 TFLOPS TF32, 148 TFLOPS FP16/BF16, and 296 TOPS INT8 — only the memory subsystem changed. We store 141 GB to match NVIDIA's usable-capacity convention for HBM3e parts (the H200 is likewise 141GB usable on 144GB physical); press reports describe the announced part as 144 GB HBM3e, so 141 vs 144 is a usable-vs-physical distinction worth reviewing. Memory bandwidth is carried at 4.0 TB/s (the confirmed H20-family figure); NVLink is 900 GB/s. FP8 is omitted to stay consistent with the existing H20 96GB row, which also omits it.
Performance
Peak theoretical throughput by precision type
| Precision | Dense | 2:4 Sparse |
|---|---|---|
FP64 64-bit floating point | 1TFLOPS | Structured sparsity is a tensor-core feature; this vector precision has no sparse form |
FP32 32-bit floating point | 44TFLOPS | Structured sparsity is a tensor-core feature; this vector precision has no sparse form |
TF32 TensorFloat-32 | 74TFLOPS | The part supports 2:4 sparsity, but the vendor publishes no sparse figure for this precision |
BF16 Brain Float 16 | 148TFLOPS | The part supports 2:4 sparsity, but the vendor publishes no sparse figure for this precision |
FP16 16-bit floating point | 148TFLOPS | The part supports 2:4 sparsity, but the vendor publishes no sparse figure for this precision |
FP8 8-bit floating point | 296TFLOPS | 592TFLOPS |
FP6 | No verified data available | The part supports 2:4 sparsity, but the vendor publishes no sparse figure for this precision |
FP4 | No verified data available | The part supports 2:4 sparsity, but the vendor publishes no sparse figure for this precision |
INT8 8-bit integer | 296TOPS | The part supports 2:4 sparsity, but the vendor publishes no sparse figure for this precision |
Dense peak divided by accelerator TDP. Board power only: excludes host CPUs, networking, cooling and facility overhead. TDP for this part: 400 W.
The H20 141GB HBM3e in the GPU landscape
Peak FP16 TFLOPS (dense) against TDP, single-GPU parts tracked by Flopper
Higher and further left is better: more half-precision throughput for less power.
Specifications
Architecture
Hopper
Form Factor
SXM
Launch Year
2025
Process Node
4nm
Memory
141 GB HBM3e
Bandwidth
4,000 GB/s
TDP
400 W
Max power (Flopper estimate)
~460 W est. Flopper estimate: 400 W TDP x 1.15. The vendor publishes no maximum board power for this part.
Interconnect
900 GB/s NVLink
bidirectional, per GPU
Spec Confidence
Vendor claimed
Full Specifications
| Memory | |
|---|---|
| Memory | 141 GB |
| Memory Type | HBM3e |
| Bandwidth | 4.0 TB/s |
| Interface Width | No verified data available |
| Interconnect & I/O | |
| GPU-to-GPU | NVLink |
| Interconnect Bandwidth | 900 GB/s bidirectional, per GPU |
| Power & Thermal | |
| TDP | 400 W |
| Max power (Flopper estimate) | ~460 W est. |
| Enterprise Features | |
| Sparsity | Yes |
| General | |
| Form Factor | SXM |
| Architecture | Hopper |
| Process Node | 4nm |
| Launch Year | 2025 |
Datasheet & Resources
Flopper Datasheet
NVIDIA H20 141GB HBM3e specifications, generated from our database. Printable.
Implications of the H20 coming back (H20 / H20E HBM3e analysis)
TechZephyr · 2025-07-19
NVIDIA Developer Documentation
Technical specs, programming guides
Data Provenance
Every figure traced to a source
Primary Source
- Publisher
- TechZephyr
- Published
- 2025-07-19
Data Quality
- Spec confidence
- Vendor claimed
- Clock basis
- Boost
- Core precisions with figures
- 7 of 9
- Normalization
- All values in TFLOPS
Compare cloud providers offering on-demand GPU instances for AI training, inference, and HPC workloads.
Browse GPU Cloud ProvidersSimilar GPUs
Frequently Asked Questions
How many TFLOPS does the NVIDIA H20 have?
The NVIDIA H20 delivers 44 TFLOPS FP32, 148 TFLOPS FP16 and 296 TFLOPS FP8 at peak.
What is the power consumption of the NVIDIA H20?
The NVIDIA H20 has a TDP (Thermal Design Power) rating of 400 watts.
How much memory does the NVIDIA H20 have?
The NVIDIA H20 is equipped with 141 GB of memory with 4,000 GB/s of memory bandwidth.
What architecture is the NVIDIA H20 based on?
The NVIDIA H20 is based on the Hopper architecture, launched in 2025.
Stay Updated on GPU Releases
Get notified when new GPUs are added or specifications are updated.
No spam, unsubscribe anytime.