Although we see NVIDIA aiming towards Blackwell at full throttle, the old-gen GPU Hopper (a cash cow) still receives additional updates. Most recently, NVIDIA has released the Hopper H200 NVL PCIe GPU for traditional rack server connections, suitable for general organizations that are not large-scale tech companies. NVIDIA states that over 70% of larger organizations’ servers are typically rack-mounted with power not exceeding 20kW and dissipate heat through air, making the original H200 model with NVIDIA’s own SXM socket unsuitable. To cater to this market, the company has introduced the PCIe slot H200 model, utilizing a high-speed NVLink GPU interconnect protocol (supporting up to 4 GPUs) to maintain data transmission performance (although performance numbers may slightly decrease compared to the SXM model). Customers of the NVIDIA H200 NVL will also receive a 5-year license for NVIDIA AI Enterprise software, a perk not included in the SXM version. Source – NVIDIA.
TLDR: Despite NVIDIA’s focus on Blackwell, the latest Hopper H200 NVL PCIe GPU is designed for traditional rack server connections in organizations, offering a 5-year AI software license and high-speed data transmission through NVLink.
Leave a Comment