Nvidia, along with its partners, has commenced the large-scale manufacturing of Blackwell GPUs designed for AI and high-performance computing (HPC), as well as the servers that incorporate these GPUs. This announcement was made by Jensen Huang, the CEO of Nvidia, during CES. He confirmed that leading cloud service providers have already implemented Blackwell systems, and there are configurations available for global data center integration.
“Blackwell has entered full production,” Huang stated. “It’s quite impressive to see. Every major cloud service provider now has operational systems.”
Although the Blackwell GPUs elevate the computational power and efficiency over the previous Hopper-generation processors for AI and HPC tasks, they do require more energy, presenting challenges for data centers that need enhanced cooling and power infrastructure. For instance, a data rack that needed 40 kW with Hopper-based systems might now need up to 120 kW with a Blackwell-equipped 72 GPU setup.
Dell was ahead in delivering Blackwell-based units, starting shipments in mid-November to certain cloud service providers. However, Dell isn’t the only one in this market. Nvidia has noted that there are over 200 unique configurations from more than a dozen server manufacturers, ensuring a broad availability of Blackwell-based systems suitable for various data centers across the globe.
“We have offerings from around 15 computer manufacturers in roughly 200 different SKUs,” Huang explained. “These include systems that are liquid-cooled, air-cooled, feature x86 architecture, Nvidia’s own Grace CPU, NVL36×2, and NVL72×1 configurations. There’s a wide variety of systems to meet the needs of virtually every data center worldwide.”
According to Nvidia’s CEO, these systems are now being mass-produced. Contrary to earlier reports that suggested the discontinuation of dual-rack 72-way GB200-based NVL36×2 systems due to lack of substantial benefits, it appears that some manufacturers are either still producing or planning to produce these dual-rack systems.
“These systems are manufactured in about 45 factories, indicating the widespread adoption of artificial intelligence and the industry’s rapid embrace of this new computing paradigm,” added Huang.
