H100 NVL GPU

The H100 NVL is NVIDIA's Hopper architecture GPU designed for data center deployment, serving as a more power-efficient variant of the H100 line with reduced power consumption compared to the SXM model. Built on the Hopper architecture, it occupies a position in NVIDIA's previous-generation lineup, having been succeeded by the GB300 Blackwell Ultra series in the company's product hierarchy. The NVL variant features 94 GB of HBM3 memory and incorporates fourth-generation Tensor Cores alongside the Transformer Engine for AI workloads. The H100 NVL delivers 835 TFLOPS of FP16 performance through its 14,592 CUDA cores and 456 Tensor cores, with 3,958 GB/s of memory bandwidth. Its 400W TDP makes it more power-efficient than higher-wattage variants while maintaining substantial compute capability. The GPU supports Multi-Instance GPU (MIG) technology for workload isolation and includes built-in Confidential Computing capabilities for secure processing scenarios. In cloud deployments, the H100 NVL finds application in large language model training and inference workloads where its substantial VRAM and Transformer Engine optimization provide benefits. The lower power draw compared to SXM variants makes it suitable for deployments where thermal and power constraints are considerations, while still delivering performance for AI training, high-performance computing applications, and accelerated data analytics tasks.