HPE has its first order for AMD's Helios rack-scale AI system, and it is a sizable one: $1.2 billion from cloud operator Vultr.
Vultr, which HPE describes as the world's largest privately held cloud infrastructure company, intends to fill space in its U.S. data centers with racks marketed as the AMD Helios AI Rack by HPE. In HPE's account, every rack carries 72 AMD Instinct MI455X GPUs, paired with EPYC "Venice" processors, Pensando Vulcano network cards and the ROCm software stack. AMD designed that side of the system for training very large models and running inference at volume. HPE's share is six QFX5252 switch trays per rack that tie all 72 GPUs together over Ethernet built on open standards, plus liquid cooling and deployment services.
The switching also ties the order to HPE's Juniper business. Scale-up means the links among GPUs inside one rack, as opposed to the network joining racks to each other, and that is where the Ethernet trays sit. HPE adds that the order extends nearly three years of joint effort between Vultr and Juniper Networks, now part of HPE, and that the rack supports open fabrics such as UALink over Ethernet. Since AMD designed the compute, the switch trays are the hardware HPE brings itself to this system.
The announcement leaves out the figures that would size the deal: how many racks the $1.2 billion covers, when they ship, which U.S. sites will host them, and how the total divides among compute, networking and services. Without a rack count, nobody can work out what a single rack costs.
HPE chief executive Antonio Neri says the order shows the company can "engineer the complete AI architecture," from rack-scale compute to networking, cooling and services. Vultr chief executive J.J. Kardwell says demand "continues to outpace available capacity," and that the partnership lets Vultr bring Helios compute and scale-up networking online sooner for its customers.
