Ampere unveils 512-core AmpereOne Aurora chip with integrated AI acceleration

Arm server chip upstart Ampere Computing took the wraps off its next generation of products, which will reach 512 cores in the next few years and offer an AI call processing unit.

But don’t be planning to place an order just yet. This next generation processor, dubbed Aurora, is not due until 2026. For now, the AmpereOne currently shipping has 192 cores, with the 256-core AmpereOne MX due next year. Ampere’s AI processor will also be a part of Aurora.

“For the first time, using our own Ampere AI IP that we integrate directly into the SOC via our interconnect [and] also high bandwidth memory attached to this platform, we are really addressing those key critical AI use cases, starting with inference, but also scaling into training as well,” said Jeff Wittich, chief product officer at Ampere Computing, on a conference call with the press.

Aurora also comes with a scalable AmpereOne Mesh, which the company claims allows for the seamless connection of all types of compute and a distributed coherence engine that supports coherency across all nodes. Ampere claims Aurora will deliver three times the performance per rack compared to the current flagship AmpereOne processors.

Aurora offers powerful AI compute capabilities for workloads like RAG and vector databases, but Wittich said it will support all types of enterprise applications, not just cloud. “So it is easy to deploy everywhere, not just hyperscalers,” he said.

Wittich also noted that Aurora can be air-cooled, making it deployable in any existing data center without requiring a retrofit to support liquid cooling. He emphasized the AmpereOne product line’s power efficiency, saying it is more power-friendly for existing data centers.

“The fact is that 77% of all data centers in the world had a maximum per rack power draw of less than 20 kilowatts, and more than half the racks out there are less than 10 kilowatts. So that means that those really big solutions, like the Nvidia DGX box, can’t even go into over half of today’s data centers,” he said.

“So, the vast majority of data centers really need efficient solutions that fit into their existing air-cooled environments. Otherwise, AI is only going to work in a few geographies with a few companies,” Wittich said.

Read more CPU news

• Intel highlights new Xeons for AI at Hot Chips 2024
• Nvidia fixes chip glitch, says production back on schedule for Q4
• 3 takeaways from AMD’s ZT Systems grab
• AWS launches Graviton4 instances
• F5 teams with Intel to boost AI delivery, security
• Everyone but Nvidia joins forces for new AI interconnect