Nvidia H800
H100 variant with reduced interconnect for the Chinese market.
Last update 3 hours ago
Key Specifications
- Architecture
- Hopper
- Memory per GPU
- 80 GB HBM2e
- Memory bandwidth
- 2,000 GB/s
- GPU interconnect
- NVLink 400 GB/s
- Release date
- Q1 2023
We couldn't find any available H800 GPUs. Search for alternative GPUs.
Technical Specifications
Nvidia H800 · Per GPU
| Compute · dense | |
|---|---|
| FP8 | 1,513 TFLOPS 3,026 with sparsity |
| FP16 / BF16 | 756.5 TFLOPS 1,513 with sparsity |
| INT8 | 1,513 TOPS 3,026 with sparsity |
| FP32 | 51 TFLOPS |
| Precision support | FP8FP16BF16TF32FP32FP64INT8 |
| Memory | |
|---|---|
| Capacity | 80 GB HBM2e |
| Bandwidth | 2,000 GB/s |
| Bus width | 5,120-bit |
| ECC | Yes |
| Silicon | |
|---|---|
| Architecture | Hopper |
| Process | TSMC 4N |
| Transistors | 80 billion |
| Shader cores | 14,592 CUDA cores |
| Matrix cores | 456 Tensor cores |
| Compute units | 114 SMs |
| Fabric and host | |
|---|---|
| GPU interconnect | NVLink 400 GB/s |
| Host interface | PCIe 5.0 x16 |
| Power | |
|---|---|
| Board power | 350 W |
| Platform | |
|---|---|
| Partitioning | MIG, up to 7 instances |
Frequently Asked Questions
Why choose the H800?
-
80GB HBM3 with FP8 Transformer Engine, similar compute to H100. Available in the Chinese market where H100 is restricted. 400 GB/s NVLink still supports multi-GPU training.
When is the H800 not a good fit?
-
Reduced NVLink bandwidth (400 vs 900 GB/s) compared to H100. Not ideal for communication-heavy distributed training across many nodes. Outside China, the H100 is typically a better choice.
What data types does the H800 support?
-
The H800 supports 7 precision formats. Training: BF16, FP16, TF32, FP32. Inference: FP8, INT8. Scientific: FP64.
Does the H800 support NVLink?
-
Yes. 400 GB/s bidirectional.
Alternatives to Nvidia H800
Last updated