Nvidia says Blackwell-based servers are in full production - 200 different configurations now available

Nvidia and its partners have initiated full-scale production of Blackwell GPUs for AI and HPC, as well as servers based on them, announced Jensen Huang, chief executive of the company, at CES. All the major cloud service providers now have Blackwell systems up and running, and Nvidia's partners have systems that can fit into all data centers worldwide. 

"Blackwell is in full production," said Huang. "It is incredible what it looks like, so first of all […] every single cloud service provider now has systems up and running." 

While Nvidia's Blackwell GPUs for AI and HPC applications significantly increase compute performance and compute performance per watt compared to Hopper-generation processors, they also consume significantly more power. This makes installing them into data centers harder as they require more cooling and power. If a Hopper-based rack consumes 40 kW, then a Blackwell-based rack with 72 GPUs reportedly consumes up to 120 kW. 

Dell was the first company to start shipments of Blackwell-based machines in mid-November to select cloud service providers, but it is not the only company to offer such servers today. Nvidia says that with over 200 different configurations from over a dozen server makers, so there are now Blackwell-based systems that can fit into a wide range of data centers. 

"We have systems here from about 15 computer makers it is being made in about 200 different SKUs, 200 different configurations," Huang said. "There are liquid cooled, air cooled, x86, Nvidia Grace CPU versions, NVL36×2, NVL72×1. A whole bunch of different types of systems so that we can accommodate just about every single data center in the world well." 

These machines are mass-produced today, according to the head of Nvidia. Interestingly, earlier reports claimed that Nvidia had canceled dual-rack 72-way GB200-based NVL36×2 systems as they did not offer compelling value and chose to focus on the single-rack NVL72 and NVL36 offerings. Apparently, this is not the case, and some companies either produce dual-rack NVL36×2 systems today or plan to do so in the future. 

"These systems are being manufactured in some 45 factories, which tells you how pervasive artificial intelligence is and how much the industry is jumping onto artificial intelligence in this new computing model," Huang said.