The H100 NVL is the high-memory PCIe variant of Hopper: 94GB of HBM3 on a 6,016-bit interface with 3.9 TB/s of bandwidth.
Designed for LLM inference at scale, it pairs over a 600 GB/s NVLink bridge for 188GB of combined memory.
Technical specifications
Compute
GPU model
NVIDIA H100 NVL
Architecture
NVIDIA Hopper
Memory
GPU memory
94GB HBM3
Memory interface
6,016-bit
Memory bandwidth
3.9 TB/s
System
Interface
PCIe Gen5 x16 (128 GB/s)
NVLink
600 GB/s with NVLink bridge
Multi-Instance GPU
Up to 7 MIG instances @ 12GB
Power
350–400W, configurable
Form factor
Dual-slot, full-height, full-length
Cooling
Passive — server airflow required
Notes
Passively cooled data center card — requires a server chassis with front-to-back airflow.
Price is per card. Multi-unit quotes available on request.
Compliance & Important Information
Availability, pricing, and specifications are subject to change without notice and seller confirmation. Equipment is offered subject to prior sale. Unless otherwise stated, all equipment is sold as-is, where-is, without warranty or guarantee. Buyers are responsible for conducting their own inspection and technical verification prior to purchase. All transactions are subject to applicable export controls, sanctions, and other legal and compliance requirements.