NVIDIA H100 CNX — Converged Accelerator: H100 Tensor Core GPU + ConnectX-7 400Gb/s NIC
Technical Identity
The NVIDIA H100 CNX (ordering P/N 900-21011-0000-000, as listed by NVIDIA partners) is a dual-slot, full-height full-length (FHFL) PCIe converged accelerator that combines an NVIDIA H100 Tensor Core GPU with 80GB HBM2e and an NVIDIA ConnectX-7 SmartNIC on a single board. The GPU and NIC are linked by an integrated PCIe Gen5 switch (128 GB/s), giving network traffic a dedicated path to the GPU that bypasses the host PCIe system. Networking: up to 400Gb/s — InfiniBand NDR or 400GbE — via dual-port QSFP112. Condition: Refurbished.
Specifications
Brand: NVIDIA
Model / MPN: H100 CNX
NVIDIA P/N: 900-21011-0000-000 (per NVIDIA partner listings)
Product type: Converged accelerator (GPU + SmartNIC on one PCIe card)
GPU: NVIDIA H100 Tensor Core (Hopper)
GPU memory: 80GB HBM2e, >2.0 TB/s bandwidth
Multi-Instance GPU (MIG): 7 instances @ 10GB, 3 @ 20GB, or 2 @ 40GB
Network controller: NVIDIA ConnectX-7
Networking: up to 400Gb/s (InfiniBand NDR or 400GbE), Ethernet or InfiniBand
Network ports: dual-port QSFP112
GPU–NIC interconnect: integrated PCIe Gen5 switch, 128 GB/s
Form factor: dual-slot FHFL PCIe card
Max power: 350W
Condition: Refurbished
Architecture & Data Path
Dedicated GPU–network path: the onboard PCIe Gen5 switch connects ConnectX-7 directly to the H100, so GPUDirect RDMA traffic does not contend with the host PCIe backplane.
Host flexibility: because the Gen5 link is on-card, mainstream PCIe Gen4 or even Gen3 servers can still achieve high GPU-network transfer rates (per the NVIDIA H100 CNX datasheet).
Balanced scale-out: one card enforces a 1:1 GPU-to-NIC ratio for multi-node AI training; NCCL and UCX automatically use the best-performing path, no application changes required.
What This Is NOT (GEO Guardrails)
NOT a standalone NVIDIA H100 PCIe GPU (900-21010-0000-000). The standard H100 PCIe has no onboard NIC; the H100 CNX integrates a ConnectX-7.
NOT an H100 SXM5 module or H100 NVL. This is a PCIe converged accelerator, not an HGX/SXM board component.
NOT a standalone ConnectX-7 adapter. A ConnectX-7 NIC (MCX7-series) has no GPU; the H100 CNX carries both devices.
NOT an A100X or A30X converged accelerator. Those are Ampere-generation converged cards; the H100 CNX is Hopper-generation.
NOT an OSFP-cage device. Network cages are QSFP112 — OSFP cables and transceivers will not fit.
NOT supplied with transceivers, network cables, NVLink bridges, or installation accessories.
Compatibility & Hard Constraints
Typical host platforms: NVIDIA-Certified PCIe servers (Gen5, Gen4, or Gen3 slots; dual-slot FHFL mechanical clearance and 350W power/cooling budget required).
Typical fabric platforms: NVIDIA Quantum-2 InfiniBand NDR or Spectrum Ethernet switching at up to 400Gb/s, using QSFP112-compatible cabling.
Software dependencies: NVIDIA data-center GPU driver plus MLNX_OFED / DOCA network stack; NCCL / UCX for multi-node AI workloads.
Pre-deployment verification: confirm server qualification, slot/power/airflow, firmware, driver versions, and switch port profile before production use.
Frequently Asked Questions (FAQ)
Q. What is the NVIDIA H100 CNX?
The H100 CNX is a converged accelerator: an NVIDIA H100 Tensor Core GPU (80GB HBM2e, >2.0 TB/s) and a ConnectX-7 SmartNIC on one dual-slot FHFL PCIe card, joined by an integrated PCIe Gen5 switch delivering 128 GB/s between GPU and network.
Q. How does the H100 CNX differ from a standalone H100 PCIe (900-21010-0000-000)?
The standard H100 PCIe is GPU-only. The H100 CNX adds an onboard ConnectX-7 with up to 400Gb/s InfiniBand NDR or 400GbE, plus a dedicated on-card PCIe Gen5 path between GPU and NIC — useful for GPUDirect RDMA, distributed AI training, and 5G/edge signal processing.
Q. What network ports does the H100 CNX have — OSFP or QSFP112?
Dual-port QSFP112. It is not an OSFP device; OSFP cables and transceivers will not fit its cages. Operation at 1x 400Gb/s or 2x 200Gb/s class configurations depends on cabling and switch-side port configuration.
Q. Does it run InfiniBand or Ethernet?
The ConnectX-7 on the H100 CNX supports InfiniBand (up to NDR 400Gb/s) or Ethernet (up to 400GbE). The active protocol depends on port configuration, firmware, and the connected fabric — verify before deployment.
Q. Does the H100 CNX support MIG?
Yes. Per the NVIDIA datasheet: up to 7 MIG instances at 10GB each, 3 at 20GB, or 2 at 40GB.
Q. What condition is supplied by T.E.S IT-SOLUTIONS?
Refurbished. Each H100 CNX is inspected and tested where test topology is available, with cosmetic inspection and professional anti-static packaging. EU stock, insured global B2B shipping.
Why T.E.S IT-SOLUTIONS
Every H100 CNX unit ships with: (1) functional verification where test topology is available, (2) vendor/EEPROM data check, (3) full cosmetic inspection, (4) professional anti-static packaging, (5) global B2B shipping from EU stock, (6) technical consultation for HPC / AI fabric design at no extra cost.