Graphics Card PNY NVIDIA H100 NVL 94GB HBM3 PCIe

PNY NVIDIA H100 NVL PCIe 94 GB HBM3 - a Hopper accelerator for large language model inference

The H100 NVL carries 94 GB of HBM3 memory with 3938 GB/s of bandwidth - the largest capacity and the highest memory speed among the PCIe cards of the H100 family. This configuration is intended for models that do not fit into 80-gigabyte accelerators: deployment and fine-tuning of LLMs of the Llama 70B class, generative AI, processing of large context windows without splitting the model across cards.

The GH100 graphics processor built on the NVIDIA Hopper architecture (TSMC 4N process) has 14,592 CUDA cores and 456 fourth-generation Tensor cores with Transformer Engine and FP8 support - up to 3341 TFLOPS FP8 and 3341 TOPS INT8 with Sparsity. A pair of cards is joined by three NVLink bridges with up to 600 GB/s of bandwidth, forming an H100 NVL configuration with 188 GB of combined memory.

• Memory: 94 GB HBM3 with ECC, 3938 GB/s, 6016-bit bus
• 14,592 CUDA cores, 456 fourth-generation Tensor cores; FP64 - 30 TFLOPS, FP64 Tensor - 60 TFLOPS
• PCI Express Gen5 x16 interface (128 GB/s), compatible with Gen4 x16
• NVLink via 3 bridges - up to 600 GB/s between two cards
• Multi-Instance GPU: up to 7 isolated instances
• FHFL, dual-slot, passive cooling, 16-pin (12V-2x6) power, 350–400 W

The card comes without display outputs (headless) and requires system airflow in a server chassis. An NVIDIA AI Enterprise license is included; the 7x NVDEC and 7x JPEG hardware decoders speed up data preparation. We supply it as an official PNY/NVIDIA partner - distributor warranty, VAT invoice and quotation of multi-GPU and NVLink configurations for your platform. Engineer consultation: +380 93,594 00 77.