The NVIDIA H100 PCIe is the PCIe form factor variant of the Hopper architecture flagship, announced at GTC March 2022 alongside the H100 SXM. It brings Hopper capabilities to standard server infrastructure without requiring specialized SXM baseboards, making it accessible to a broader range of enterprise deployments.
The H100 PCIe uses a cut-down version of the GH100 die with 80GB of HBM2e memory (not HBM3) providing 2 TB/s of bandwidth. While lower than the SXM variant's 3.35 TB/s, this is still a significant upgrade from A100 PCIe. The chip includes 14,592 CUDA cores and 456 Tensor Cores, slightly reduced from the SXM variant.
The dual-slot PCIe Gen5 x16 form factor is compatible with standard server platforms without specialized baseboards. TDP is 350W, exactly half the SXM variant, enabling air-cooled deployments. For multi-GPU configurations, it supports NVLink bridge connecting two cards with 600 GB/s bandwidth, or PCIe peer-to-peer communication.
The H100 PCIe includes all Hopper architectural features: the Transformer Engine with FP8 precision, 4th-generation Tensor Cores, DPX instructions, and MIG support for up to 7 instances. It's particularly suited for inference workloads, smaller-scale training, and deployments where SXM infrastructure is not available or practical.