NVIDIA, according to analyst Ming-Chi Kuo, has decided to abandon the production of two-rack NVL36×2 AI systems based on GB200 accelerators in favor of single-rack NVL72 and NVL36 machines. This is explained by limited resources and customer preferences.
Initially, it was planned to release three GB200 supersystems based on Blackwell accelerators for AI and HPC workloads – NVL72, NVL36 and NVL36×2. The first combines 18 1U nodes in one rack, each of which contains two GB200 accelerators. In total, this gives 72 B200 chips and 36 Grace processors. The NVLink 5 bus is enabled, and the system power consumption is at 120 kW. In turn, NVL36 has 36 B200 chips, while NVL36×2 combines two such systems.
Image source: NVIDIA
It was expected that the NVL36×2 configuration would become more widespread than the NVL72. The fact is that the data centers of most NVIDIA customers cannot meet the NVL72 requirements in terms of power and cooling. From this point of view, NVL36×2 represents a compromise solution. On the other hand, the NVL72 requires less installation space and has lower total power consumption: each of the NVL36x2 racks requires 66 kW, for a total of 132 kW. This results in slightly lower performance.
According to Ming-Chi Kuo, some customers, in particular Microsoft, preferred NVL72 over NVL36x2. At the same time, disagreements arose between the companies regarding the configuration of the racks even at the discussion stage. In addition, managing three different projects to create super accelerators based on the GB200 became a difficult task for NVIDIA. Therefore, it was decided to abandon the two-post machine.
It is also noted that mass production of NVL72 may be delayed until the second half of 2025, although the first half of next year was previously called. However, some customers, including Microsoft, will begin receiving these systems in December.