Emerging leader in ARM server processors, Ampere Computing, recently unveiled its upcoming range of chips which is anticipated to scale up to 512 cores in the ensuing years, inclusive of a dedicated AI call processing unit.
Yet, these cutting-edge Aurora processors aren’t available for immediate purchase, with a launch expected not before 2026. Presently, Ampere Computing is distributing its AmpereOne processor that features 192 cores, and next year, it plans to release the AmpereOne MX with 256 cores. The forthcoming Aurora chips will also incorporate the company’s AI processors.
“Utilizing our proprietary Ampere AI IP, which is directly integrated into the SOC using our interconnect alongside high bandwidth memory within this platform, we are strategically targeting essential AI applications, initially focusing on inference and expanding into training,” explained Jeff Wittich, the Chief Product Officer at Ampere Computing, during a media teleconference.
The Aurora processor will also feature the scalable AmpereOne Mesh, which according to the company, facilitates a seamless integration across various computing types, and includes a distributed coherence engine that ensures coherency across all nodes. Ampere boasts that Aurora will offer a tripled performance per rack in comparison to its existing top-tier AmpereOne processors.
Aurora delivers robust AI processing power capable of handling tasks such as RAG and vector databases. However, Wittich emphasized that its utility spans beyond just cloud applications, ensuring ease of deployment across various venues, not limited to large-scale cloud providers. “So it is easy to deploy everywhere, not just hyperscalers,” he remarked.
Additionally, Wittich highlighted that Aurora hardware can function in air-cooled settings, allowing it to integrate into any standard data center without the need for adaptations for liquid cooling. He pointed out the energy efficiency of the AmpereOne series, noting its suitability for typical data center operations.
“The reality is, 77% of all data centers globally operate with a maximum per rack power consumption of under 20 kilowatts, with more than half of these setups supporting less than 10 kilowatts. This essentially rules out the deployment of larger units, like the Nvidia DGX box, in more than half of the current data centers,” he stated.
Wittich argued that the vast number of data centers require solutions that are not only efficient but also compatible with their existing air-cooled frameworks. Without such solutions, AI technology would remain limited to specific regions and a handful of corporations,” he concluded.