Product Overview
The NVIDIA H200 NVL is the PCIe form factor version of the H200, using a dual-slot passive cooling design with the same Hopper GH100 die and 141GB HBM3e memory. Unlike the H200 SXM (700W, HGX board), the H200 NVL ships as a PCIe add-in card with TDP configurable up to 600W, and delivers 900 GB/s interconnect via 2-4-way NVLink Bridges.
Strategic position: The H200 NVL targets single-card or small-scale (2-4 card) H200 deployments in standard PCIe servers, combining large memory (141GB) with air-cooling compatibility — the H200 SXM alternative for general-purpose servers.
Core Specifications
| Parameter | Value |
|---|
| Architecture | NVIDIA Hopper (GH100) |
| Process | TSMC 4nm |
| FP8 Compute | 3,341 TFLOPS |
| FP16 Compute | 1,671 TFLOPS |
| FP32 | 60 TFLOPS |
| INT8 Compute | 3,341 TOPS |
| Memory | 141 GB HBM3e |
| Memory Type | HBM3e |
| Memory Bandwidth | 4.8 TB/s |
| Interconnect | 2-4-way NVLink Bridge 900 GB/s (per GPU) |
| PCIe | Gen 5 x16 |
| TDP | Up to 600 W (configurable) |
| Form Factor | Dual-slot PCIe passive cooling |
| Release | 2024 |
| Price | approx. $38,000 - $40,000 (market average) |
Comparison with H200 SXM
| Metric | H200 NVL | H200 SXM | Difference |
|---|
| Memory | 141GB HBM3e | 141GB HBM3e | Identical |
| Bandwidth | 4.8 TB/s | 4.8 TB/s | Identical |
| FP8 | 3,341 TFLOPS | 3,958 TFLOPS | NVL 16% lower |
| TDP | 600 W | 700 W | NVL 100W lower |
| Form factor | PCIe card | SXM5 module | Different |
| Multi-card scaling | 2-4-way Bridge | Full NVSwitch 8-card | NVL smaller scale |
Key Features
- 141GB HBM3e large memory: suited to long-context LLM inference and large-batch processing
- Air-cooling compatible: deploys in standard PCIe servers
- NVLink Bridge: 900 GB/s interconnect maintained in 2-4 card configurations
Use Cases
- ✅ LLM inference (single card / small clusters)
- ✅ Long-context processing
- ✅ AI acceleration in standard PCIe servers
- ⚠️ Hyperscale distributed training (no match for SXM + NVSwitch)
External Links