← Back to comparison
NVIDIA
NVIDIA H100 PCIe 80GB
Hopper · H-Series · ACTIVE
We earn commissions from purchases made through links on this site. This doesn't affect our rankings or recommendations.
Prices updated 5 days ago
New price
—
Used avg (eBay 30d)
—
$/GB VRAM (best)
—
MSRP: $30,000
Specifications
| VRAM | 80 GB |
| Memory type | HBM3 |
| Bus width | 5120-bit |
| Memory bandwidth | 3,350 GB/s |
| CUDA cores | 14,592 |
| Tensor cores | 456 |
| FP16 | 204.9 TFLOPS |
| FP8 | 1513 TFLOPS |
| TDP | 350W |
| Power connector | 8-pin |
| Card length | 267 mm |
| Slot width | 2 slots |
| PCIe | Gen 5 x16 |
| CUDA compute | 9.0 |
| Max model (Q4) | ~156B parameters |
Inference Benchmarks (Q4_K_M)
Llama 3.3 8B
180.0 tok/s
Qwen 3 32B
75.0 tok/s
Llama 3.3 70B
35.0 tok/s
llama.cpp, batch_size=1, ctx=4096, single GPU.
What Can You Run?
| Model | Q4_K_M | Q8_0 | FP16 |
|---|---|---|---|
| Llama 3.3 8B8B | Excellent~180 tok/s | Usable | Usable |
| Llama 3.3 70B70.6B | Good~35 tok/s | Usable | Won't fit |
| Qwen 3 8B8.2B | Usable | Usable | Usable |
| Qwen 3 32B32.8B | Excellent~75 tok/s | Usable | Usable |
| DeepSeek R1 70B70.6B | Usable | Usable | Won't fit |
| Mistral Nemo 12B12.2B | Usable | Usable | Usable |
| Phi-4 14B14B | Usable | Usable | Usable |
| Gemma 3 27B27.4B | Usable | Usable | Usable |
| Codestral 25B25.3B | Usable | Usable | Usable |
| Command R 35B35B | Usable | Usable | Usable |
Notes
The datacenter gold standard. FP8 performance is exceptional.