NVIDIA Tesla H100 Graphic Card – 80 GB – Full-height 900-21010-0000-000

Description

NVIDIA H100 Tensor Core GPU 

The NVIDIA H100 Tensor Core GPU is the flagship data-center accelerator built on NVIDIA’s Hopper architecture, designed to power the most demanding AI training, AI inference, and high-performance computing (HPC) workloads in the world. While many retailers and resellers still refer to it informally as the “Tesla H100” — a holdover from NVIDIA’s older Tesla branding for data-center GPUs — the official product name is simply the NVIDIA H100 Tensor Core GPU. It has no consumer-facing “GeForce” or “RTX” branding; it’s purpose-built exclusively for servers, AI clusters, and scientific computing infrastructure.

At the heart of the H100 is the GH100 die, fabricated on a custom TSMC 4N process and packing roughly 80 billion transistors — one of the largest GPU dies ever produced at volume. Depending on form factor, the H100 enables either 132 streaming multiprocessors (SXM5 and NVL variants) or 114 streaming multiprocessors (PCIe variant), translating to up to 16,896 CUDA cores and 528 fourth-generation Tensor Cores on the SXM5 model, or 14,592 CUDA cores and 456 Tensor Cores on the PCIe model. Every H100 ships with 80GB of high-bandwidth memory — HBM3 on the SXM5 version (delivering roughly 3.35TB/s of bandwidth) or HBM2e on the PCIe version (around 2TB/s).

What truly sets the H100 apart from previous-generation GPUs is its dedicated Transformer Engine, a piece of hardware purpose-built to accelerate transformer-based deep learning models — the architecture behind today’s large language models. The Transformer Engine intelligently and automatically switches between FP8 and FP16 precision during training and inference, delivering up to 30x faster large language model performance compared to the prior Ampere generation, without sacrificing model accuracy. Combined with fourth-generation NVLink (offering 900GB/s of GPU-to-GPU bandwidth on SXM5 systems) and the NVLink Switch System, organizations can interconnect up to 256 H100 GPUs into a single, unified accelerated computing fabric capable of tackling trillion-parameter AI models.

The H100 is available in three primary form factors to fit different deployment needs. The PCIe version slots into standard servers using a familiar PCIe Gen5 x16 interface, drawing up to 350W and prioritizing flexible, lower-power deployment. The H100 NVL variant pairs two PCIe cards together via NVLink bridges, boosting total available memory and inter-GPU bandwidth for large-scale inference workloads. The SXM5 version is the performance flagship, built for NVIDIA’s HGX and DGX platforms, supporting full NVLink connectivity and power envelopes up to 700W for maximum multi-GPU scaling in dedicated AI training clusters.

Beyond raw throughput, the H100 includes second-generation Multi-Instance GPU (MIG) technology, allowing a single physical GPU to be securely partitioned into up to seven fully isolated instances — each with guaranteed quality of service — making it ideal for multi-tenant cloud and enterprise environments. It also introduces NVIDIA Confidential Computing, providing hardware-based trusted execution environments so sensitive data and AI models can be protected even while in use, a critical requirement for healthcare, financial services, and government workloads.

Though it has since been succeeded by the H200 and the Blackwell-generation B100/B200/B300 GPUs, the H100 remains one of the most widely deployed and battle-tested AI accelerators on the market, backed by a mature CUDA software ecosystem, broad cloud availability, and proven reliability at scale.