Next-generation technology. Enterprise-scale compute.
Technical resourcesTrack your quote
AI & GPU servers / NVIDIA H200
HOPPER / HGX & NVL

NVIDIA H200.
Built around bandwidth.

Complete H200 systems for memory-intensive AI and HPC. Specify HGX SXM with NVSwitch or a PCIe NVL platform with supported bridge groups.

141 GBHBM3e per GPU
4.8 TB/sMemory bandwidth / GPU
1,128 GBAggregate in 8 GPUs
8UHGX reference chassis
CONFIGURATION BRIEF

The full system, defined.

The HGX configuration references Supermicro’s SYS-821GE-TNHR, an 8U dual-Xeon system supporting eight H200 SXM GPUs. H200 NVL is a separate PCIe option; it uses a compatible PCIe chassis such as the SYS-521GE-TNRT and a different interconnect design.

HGX reference platformSupermicro SYS-821GE-TNHR, 8U, supporting HGX H100 / H200. The NexTec proposal specifies H200 SXM.
GPU baseboard8× H200 SXM with 141 GB HBM3e each; 1,128 GB aggregate. Up to 4.8 TB/s local memory bandwidth and 900 GB/s NVLink per GPU.
GPU fabricHGX eight-GPU designs use NVLink and four NVSwitch devices. GPU fabric and external scale-out network are separate parts of the architecture.
Host computeDual 4th / 5th Gen Intel Xeon Scalable. CPU models, core count, TDP, and firmware are selected against the supported OEM list.
ECC memory32 DDR5 DIMM slots; up to 8 TB platform capacity. Actual speed depends on CPU, DIMM type and DIMMs per channel.
Boot / storage2× M.2 NVMe for boot. Up to 16 hot-swap 2.5-inch NVMe bays: 12 default, 4 optional; SATA configuration also available.
Fabric expansionEight PCIe 5.0 x16 low-profile slots, plus two FHHL slots and two optional FHHL slots. Select NIC count, GPU affinity, fabric, optics, and cabling as a complete design.
Power / facilityH200 SXM up to 700 W each: 5.6 kW across eight GPUs before the host and other components. Complete feed, PSU, cooling and rack requirements are confirmed for the ordered build.
NVL alternativeH200 NVL: 141 GB HBM3e, up to 600 W, PCIe 5.0 x16. Supported 2- or 4-GPU NVLink bridge groups; it is not the HGX eight-way switched fabric.
System acceptanceGPU and fabric health, PCIe/NVLink enumeration, memory and storage checks, sustained power/thermal behavior, representative training or inference, and documented firmware/software versions.

SXM modules are HGX platform components, not standard PCIe add-in cards. NVL and SXM cannot be substituted in the same baseboard.

A chassis reference photo does not prove the installed GPU model. The gallery separately identifies the official HGX H200 baseboard.

LET’S BUILD WHAT’S NEXT

Your workload. Our starting point.

Tell us what you’re building. We’ll help scope the infrastructure.

Request a quote