Ampere updates roadmap, heads to 256 cores

Ampere Computing’s annual update on upcoming products and milestones included an increase in core count for its Arm-based AmpereOne server chips as well as a new working group for jointly developing AI SOCs.

Ampere CEO Renee James, a former top executive at Intel, said the increasing power requirements of AI processors are simply not sustainable, and Ampere represents a better power alternative. “The current path is unsustainable,” she said in a statement. “We believe that the future datacenter infrastructure has to consider how we retrofit existing air-cooled environments with upgraded compute, as well as build environmentally sustainable new datacenters that fit the available power on the grid. That is what we enable at Ampere.”

On the processor front, the company’s top-of-the-line AmpereOne processor is going from 192 cores to 256 cores and from eight-channel memory to 12-channel memory, thus expanding memory bandwidth. Jeff Wittich, Ampere chief product officer, said the existing 192-core AmpereOne can be dropped into the 12-channel platform, making it forward compatible.

Despite the enormous number of cores, liquid cooling is not required. “None of our products on our roadmap require you to go to liquid cooling. So we’re making it easy for people to stay within their current data center infrastructure. And we’re delivering even better power efficiency to what we have today,” Wittich told me.

Ampere is claiming that in Spec_Int benchmarks, an industry standard, the 256-core AmpereOne tops the performance of AMD Genoa (96 cores) by 50% and Bergamo (128 cores) by 15%. Wittich said that Intel did not have a comparable processor in terms of core count to compare against.

For data center operators that are looking to refresh and consolidate old infrastructure to reclaim space, budget, and power, AmpereOne says it delivers up to 34% more performance per rack.

“So we’re the most efficient solution when it comes to Spec_Int rate, and you can obviously translate that into performance at a rack level as well, which AMD and Intel have started to talk more about as well,” said Wittich.

He acknowledged that the benchmark is not a real-life workload, but he said that when compared to typical web stack workloads like nginx, Redis, MySQL, and Memcached, Ampere leads against the competition. “If you put all these together and look at it at the rack level, you need a lot less space, a lot less servers, and 35% less power when you’re deploying based on AmpereOne versus AMD Genoa or AMD Bergamo,” he said.

In other news, Ampere is working with Qualcomm Technologies to scale out a joint solution featuring Ampere CPUs and Qualcomm Cloud AI100 Ultra. This solution will tackle LLM inferencing on the industry’s largest generative AI models.

Ampere also announced that Meta’s Llama 3 is now running on Ampere CPUs in Oracle Cloud. Performance data shows that running Llama 3 on the 128-core Ampere Altra CPU (the predecessor to AltraOne) with no GPU delivers the same performance as an Nvidia A10 GPU paired with an x86 CPU, while using one-third of the power, according to Ampere.

Lastly, Ampere announced the formation of a Universal Chiplet Interconnect express (UCIe) working group as part of the AI Platform Alliance it formed last year. This coalition of chip designers intends to pool their resources and talent to advance AI chip development. UCIe is designed around open silicon integration, offering an open standard across the industry to build SOC-level solutions where chiplets from various companies are integrated into an SOC.

“We believe that there’s a need for open solutions across the industry that are broadly available to everybody that aren’t walled gardens and that aren’t proprietary. So we are building these best-in-class solutions at the server level with the fast time to market and give people access to the market,” said Wittich.