Advantech SKY-TESL-H100N-94P NVIDIA H100 NVL 94GB HBM3 Passive Data Centre Accelerator with NVLink and PCIe 5.0

Large AI models and simulation codes tend to be limited by memory bandwidth as much as by raw compute. The SKY-TESL-H100N-94P is the NVIDIA H100 NVL from Advantech, a compute-optimised data centre accelerator on the Hopper architecture with 94 GB of HBM3 at 3.9 TB/s. It is a passive, dual-slot PCIe card for lower-power, air-cooled enterprise rack designs, and suits generative AI and high-performance computing (HPC).

The GPU carries 14,592 CUDA cores and 456 fourth-generation Tensor Cores, rated at 60 TFLOPS FP32 and 30 TFLOPS FP64 with fast FP64 support, and its memory carries ECC. Second-generation Multi-Instance GPU can split the SKY-TESL-H100N-94P into up to seven instances, each with partitioned hardware and guaranteed quality of service, which suits sharing one accelerator across several inference services or users. Seven JPEG decoders and seven video decoders sit on the card for image and video ingest.

The card connects over PCI Express 5.0 x16 and draws up to 400 W through a 16-pin PCIe connector. It is headless and full height, and its passive heatsink relies on the server’s airflow. Two cards can be paired with three NVLink bridges at 600 GB/s, using the optional SKY-QUAD-NVL-2A bridge for dual-slot spacing or the SKY-QUAD-NVL-3A for three-slot spacing. Software, libraries and optimised AI models come through NVIDIA’s NGC catalogue.

We supply the SKY-TESL-H100N-94P tested, and fit, configure and burn-in test it in the GPU servers we build, with NVLink bridges, power and airflow set up for the card count. For more memory per GPU, the SKY-TESL-H200N-141 is the H200 NVL with 141 GB of HBM3e, and the SKY-TESL-H100-80P is the H100 PCIe card with 80 GB of HBM2e at 350 W.