HPE has secured its first order for the AMD Helios AI rack. Vultr is paying $1.2 billion for the systems. Not entirely surprisingly, they will be deployed in U.S. data centers. Each rack contains 72 Instinct MI455X GPUs, which are interconnected via HPE Juniper Networking Ethernet switches.
This is the first order for HPE’s new Helios system. Vultr will use the racks for model training and inference. According to the company, this move is intended to meet the demand from MSPs and end users. Relying solely on Nvidia computing power no longer seems sufficient; with Helios, AMD now offers a rack-scale alternative for the first time.
Demand for AI computing remains high, as Vultr CEO J.J. Kardwell also notes. His company is still operating at capacity, meaning demand exceeds supply. With Helios, Vultr says it can make rack-scale computing power and scale-up networking available to customers more quickly.
There’s also another connection on the HPE side. Vultr has been working with Juniper Networks, which is now part of HPE, for nearly three years.
72 GPUs connected via Ethernet
AMD and HPE announced the AMD Helios rack as an Ethernet-based AI system, in contrast to the priority that Nvidia’s own InfiniBand holds in its rack systems. It follows the Open Rack Wide specifications of the Open Compute Project and utilizes liquid cooling. Each rack combines 72 AMD Instinct MI455X GPUs with Epyc Venice CPUs, Pensando Vulcano AI NICs, and the ROCm software stack. Six HPE Juniper Networking QFX5252 switch trays per rack connect all the GPUs. The design also supports open fabrics such as UALink over Ethernet (UALoE).
In some respects, the specifications outperform those of Nvidia, the market leader. AMD claims up to 2.9 exaflops of FP4 computing power and 31 TB of HBM4 memory per rack. That would represent a 50 percent increase in HBM capacity compared to Nvidia’s Vera Rubin NVL72, its direct competitor.
“AMD Helios by HPE demonstrates HPE’s unique ability to design the complete AI architecture, featuring rack-scale compute, scale-up networking, and direct liquid cooling,” says HPE CEO Antonio Neri.
AMD is attracting major customers
Vultr is one of several companies that have adopted AMD’s rack-scale solution. Meta is also set to deploy 6 gigawatts of AMD GPUs for its AI infrastructure. Prior to this, OpenAI had already signed a $90 billion deal with the chipmaker.
HPE is launching the Helios rack as part of its AI Data Center Solutions portfolio. For the deployment and liquid cooling of large AI clusters, the company offers support through HPE Services, ensuring that these complex and heavy-duty systems are deployed properly and continue to run smoothly.