Skip to main content

NVIDIA H200 NVL

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​

ParameterValue
ArchitectureNVIDIA Hopper (GH100)
ProcessTSMC 4nm
FP8 Compute3,341 TFLOPS
FP16 Compute1,671 TFLOPS
FP3260 TFLOPS
INT8 Compute3,341 TOPS
Memory141 GB HBM3e
Memory TypeHBM3e
Memory Bandwidth4.8 TB/s
Interconnect2-4-way NVLink Bridge 900 GB/s (per GPU)
PCIeGen 5 x16
TDPUp to 600 W (configurable)
Form FactorDual-slot PCIe passive cooling
Release2024
Priceapprox. $38,000 - $40,000 (market average)

Comparison with H200 SXM​

MetricH200 NVLH200 SXMDifference
Memory141GB HBM3e141GB HBM3eIdentical
Bandwidth4.8 TB/s4.8 TB/sIdentical
FP83,341 TFLOPS3,958 TFLOPSNVL 16% lower
TDP600 W700 WNVL 100W lower
Form factorPCIe cardSXM5 moduleDifferent
Multi-card scaling2-4-way BridgeFull NVSwitch 8-cardNVL 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

Vendor Information​

ParameterValue
CompanyNVIDIA Corporation
Websitehttps://www.nvidia.com
Release2024
ArchitectureHopper

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)