Ampere Computing’s annual update on future products and milestones entailed a rise in core count for its Arm-based AmpereOne server chips as well as a new working group for jointly creating AI SOCs.
Renee James, Ampere’s CEO and a previous high-ranking executive at Intel, expressed that the escalating energy demands of AI processors are simply unsustainable, and Ampere is a superior power option. She stated, “The current path is untenable. We assume that future datacenter infrastructure must consider how to upgrade existing air-cooled settings with advanced compute, and also build environmentally sustainable new datacenters that are in line with the available power on the grid. This is what we facilitate at Ampere.”
Speaking of the processor, the company’s premium AmpereOne processor is going from 192 cores to 256 cores and from eight-channel memory to 12-channel memory, thereby extending memory bandwidth. Jeff Wittich, the chief product officer at Ampere, enlightened that the current 192-core AmpereOne can be integrated into the 12-channel platform, thus making it forward compatible.
In spite of the large number of cores, liquid cooling is not a necessity. Wittich informed me, “None of our products on our roadmap require you to switch to liquid cooling. So we’re making it convenient for people to stay within their existing data center infrastructure. And we’re providing even better power efficiency than we have today.”
Ampere insists that in the industry standard Spec_Int benchmarks, their 256-core AmpereOne exceeds the performance of AMD’s Genoa (96 cores) by 50% and Bergamo (128 cores) by 15%. They mention that Intel does not have a processor with a comparable number of cores for a fair comparison.
If data center operators are seeking to renovate and consolidate outdated infrastructure to save space, budget, and power, AmpereOne promises to offer up to 34% higher performance per rack.
Wittich, a representative from Ampere, asserted that though Spec_Int rate benchmark may not reflect real-world workload, Ampere has an edge over its competition when weighed against typical web stack workloads – nginx, Redis, MySQL, and Memcached. “Assessing this at the rack level, AmpereOne requires significantly less space, fewer servers, and saves on power consumption by 35%, compared to AMD Genoa or AMD Bergamo,” Wittich stated.
In other news, Ampere is collaborating with Qualcomm Technologies to expand a combined solution featuring Ampere CPUs and Qualcomm Cloud AI100 Ultra. This solution aims to handle LLM inferencing on the industry’s most significant generative AI models.
Ampere also revealed that Meta’s Llama 3 is now operating on Ampere CPUs in Oracle Cloud. Performance data suggests that running Llama 3 on the 128-core Ampere Altra CPU (the predecessor to AltraOne) with no GPU delivers equivalent performance to an Nvidia A10 GPU paired with an x86 CPU, while using only a third of the power, as per Ampere.
In addition, Ampere announced the establishment of a Universal Chiplet Interconnect express (UCIe) working group as a part of the AI Platform Alliance it created last year. This consortium of chip designers plans to combine their resources and skills to foster AI chip development. UCIe is conceptualized around open silicon integration, providing an open standard across the industry for the creation of SOC-level solutions where chiplets from multiple companies are incorporated into an SOC.
“We believe that there’s a need for open solutions across the industry that are broadly accessible to everyone that aren’t walled gardens and that aren’t proprietary. So we are constructing these top-tier solutions at the server level with quick time to market and providing people access to the market,” said Wittich.
