Get the latest tech news
Nvidia's Blackwell Reworked – Shipment Delays and GB200A Reworked Platforms
MGX GB200A NVL36, B102, B20, CoWoS-L, CoWoS-S, GB200A NVL64, ConnectX-8, Liquid Cooling vs Air Cooling, NVLink Backplane, PCB, CCL, Substrate, BMC, Power Delivery
While Nvidia could have enabled a NVL36x2 design by keeping the switch trays with 2 NVSwitch ASICs, that would increase costs, and would make it potentially impossible to air cool due to the front OSFP NVLink cages blocking airflow. 1,500W for a 1U chassis isn’t insane on its own, but it is once you consider cooling challenges due to the Ultrapass flyover cables from the switch ASIC to the backplane connectors blocking a lot of the air flow. Unfortunately, due to the unreliability of the hardware, Nvidia recommends that at least one compute tray per NVL rack be kept in reserve to allow for GPUs to be taken offline for maintenance and thus for use as a hot spare.
Or read this on Hacker News
