The NVIDIA HGX H100 SXM5 8‑GPU Board (part number 935‑24287‑0301‑000) integrates eight NVIDIA Hopper‑based H100 SXM5 GPUs, each with 80 GB HBM3 memory, into a single high-density accelerator board. Built for AI training, inference, and exascale HPC, it supports direct liquid cooling. With NVLink at 900 GB/s interconnect and 3.35 TB/s memory bandwidth, this platform delivers unmatched scalability and performance.
| Feature |
Details |
| Part Name / Number |
NVIDIA HGX H100 SXM5 8‑GPU Board – 935‑24287‑0301‑000 |
| Form Factor |
SXM5 (8 GPUs on single board, direct liquid-cooled) |
| GPUs |
8 × NVIDIA H100 SXM5, each 80 GB HBM3 |
| Total GPU Memory |
640 GB HBM3 |
| Memory Bandwidth |
3.35 TB/s per GPU via HBM3, NVLink at 900 GB/s per NVLink
|
| CUDA Cores (Total) |
8 × 16,896 = 135,168 cores |
| Tensor Cores (Total) |
8 × 528 = 4,224 cores |
| GPU Interconnect |
NVLink SXM5 — 900 GB/s GPU ↔ GPU |
| Form Thermal Design Power |
Configurable up to 700 W per GPU |
| Cooling |
Direct liquid cooling (equipment not included) |
| Power Interface |
Board-level power connectors (server-integrated power delivery) |
| PCIe Host Interface |
PCIe Gen 5 x16 interface (board to host) |
| Multi‑Instance GPU (MIG) |
Each H100 supports up to 7 MIG instances |
| Use Cases |
Distributed AI training, supercomputing, HPC clusters |
| Certifications |
Hopper microarchitecture, NVLink-enabled, NVIDIA HGX/DGX compliant |
❓ Frequently Asked Questions (FAQs)
Q1: What is the total GPU memory on this board?
A: The board provides 640 GB total HBM3 memory—80 GB across 8 SXM5 GPUs.
Q2: How do the GPUs communicate within the board?
A: Via SXM5 NVLink, offering up to 900 GB/s per connection, enabling high-speed GPU-to-GPU communication.
Q3: What cooling solution does this board require?
A: It requires direct liquid cooling, typically implemented in purpose-built HX/L48 chassis. (Board is liquid-cooled ready.)
Q4: Is this board plug-and-play with standard servers?
A: No. It's designed for HGX-compatible host systems with integrated power, liquid cooling, and server infrastructure.
Q5: Can I run multi-instance workloads?
A: Yes—each H100 supports up to 7 MIG instances, suitable for concurrent multi-tenant or mixed workloads.