NVIDIA H100 SXM vs H100 PCIe
Compare published specifications with current rental availability.
Differences are derived only from verified memory, bandwidth, power, form factor, precision support, software ecosystem and real inventory.
| Field | NVIDIA H100 SXM 80GB | NVIDIA H100 PCIe 80GB |
|---|---|---|
| Vendor | NVIDIA | NVIDIA |
| Exact SKU | H100 SXM | H100 PCIe |
| Architecture | Hopper | Hopper |
| Generation | Hopper | Hopper |
| Category | Data center | Data center |
| Lifecycle | Current | Current |
| VRAM | 80 GB | 80 GB |
| Memory technology | HBM3 | HBM2e |
| Bandwidth | 3,350 GB/s | 2,000 GB/s |
| Power | 700 W | 350 W |
| Interface | PCIe Gen5 host interface | PCIe Gen5 x16 |
| Form factor | SXM · Module | Dual slot · PCIe · Accelerator |
| Interconnect | NVLink | PCIe |
| MIG / partitioning | Supported | Supported |
| Numeric precision formats | FP64, FP32, TF32, BF16, FP16, FP8, INT8 | FP64, FP32, TF32, BF16, FP16, FP8, INT8 |
| Runtime ecosystem | cuda, vllm, sglang, tensorrt-llm | cuda, vllm, sglang, tensorrt-llm |
| Rental availability | Available | Limited capacity |
| Monthly rental price | USD 1,860 | USD 1,579 |
Verified differences
- NVIDIA H100 SXM 80GB has the highest recorded memory bandwidth.
- NVIDIA H100 PCIe 80GB has the lowest published power value.
Architecture differences
Source-verified facts: NVIDIA H100 SXM 80GB: Hopper / Hopper / SXM · Module · NVIDIA H100 PCIe 80GB: Hopper / Hopper / Dual slot · PCIe · Accelerator
VRAM and memory requirements
Source-verified facts: NVIDIA H100 SXM 80GB: 80 GB / HBM3 / 3,350 GB/s · NVIDIA H100 PCIe 80GB: 80 GB / HBM2e / 2,000 GB/s
Power and cooling
Source-verified facts: NVIDIA H100 SXM 80GB: 700 W / SXM · Module · NVIDIA H100 PCIe 80GB: 350 W / Dual slot · PCIe · Accelerator
AI precision support
Source-verified facts: NVIDIA H100 SXM 80GB: FP64, FP32, TF32, BF16, FP16, FP8, INT8 · NVIDIA H100 PCIe 80GB: FP64, FP32, TF32, BF16, FP16, FP8, INT8
Interface and interconnect
Source-verified facts: NVIDIA H100 SXM 80GB: PCIe Gen5 host interface / NVLink · NVIDIA H100 PCIe 80GB: PCIe Gen5 x16 / PCIe
Multi-GPU considerations
Large variants may require tensor parallelism or multiple nodes. Validate total resident weights, runtime support, interconnect, topology and headroom before choosing a GPU count.
Model and workload fit
Model fit depends on exact checkpoint precision, runtime support, memory headroom and topology. Reference compatibility does not imply FusionPointAI rental inventory.
Rental availability
Source-verified facts: NVIDIA H100 SXM 80GB: Available / USD 1,860 / month · NVIDIA H100 PCIe 80GB: Limited capacity / USD 1,579 / month
Official sources
Last verified: 2026-08-25