Ampere and Qualcomm are not the most obvious partners. Both, in the end, provide Arm-based processors for data centre servers (but Qualcomm’s primary business is still mobile). But as the two firms revealed today, they are now teaming together to provide an AI-focused server that runs models rather than trains them using Qualcomm’s Cloud AI100 Ultra AI inferencing processors and Ampere’s CPUs.
Ampere wants in on the AI boom, just like every other chip maker. Although the business may leverage the Arm IP to add some of these capabilities to its chips, it is not a necessary key capability since its concentration has always been on quick and energy-efficient server processors. That is the reason Arm CTO Jeff Wittich tells me Ampere chose to collaborate with Qualcomm (and SuperMicro) to combine the two systems.
“The idea here is that while I’ll show you some great performance for Ampere CPUs running AI inferencing on just the CPUs, if you want to scale out to even bigger models—multi-100 billion parameter models, for instance—just like all the other workloads, AI isn’t one size fits all,” Wittich told Eltrys. “We’ve been working with Qualcomm on this problem; using their extremely efficient cards, we’ve got a server-level solution, combining our super-efficient Ampere CPUs to do a lot of the general-purpose tasks that you’re running in conjunction with inferencing.”
Wittich said Ampere intended to combine best-of-breed technologies when it came to cooperating with Qualcomm.
“We’ve had really excellent cooperation with Qualcomm,” he said. This is one of the projects we’ve been working on; I believe our interests are very comparable, which makes it fascinating. They are developing solutions in several market segments with very high efficiency. Our technologies on the server’s CPU side are very efficient.
Ampere included the Qualcomm collaboration in its yearly roadmap update. This plan includes the construction of a new 256-core AmpereOne chip based on cutting-edge 3nm technology. Despite Wittich’s claims of being ready at the fab and expected for release later this year, those new chips are not yet as widely accessible.
Apart from the extra cores, the 12-channel DDR5 RAM of this new generation of AmpereOne chips is what sets them apart and enables Ampere’s data centre clients to better customise their users’ memory access to suit their requirements.
Performance is just one aspect of the sales pitch; another is the cost and power consumption of running these processors in the data center. That is particularly true for AI inferencing, as Ampere enjoys benchmarking their results against A10 GPUs from Nvidia.
Noteworthy is the fact that Ampere is not retiring any of its current chips in favour of these new ones. Wittich made it clear that there are still many applications for these older processors.
Today, Ampere also disclosed another collaboration. The company is working with NETINT to develop a collaborative solution that combines NETINT‘s video processing processors with Ampere’s CPUs. This new server will concurrently subtitle 40 streams and transcode 360 live video channels using the Whisper speech-to-text model from OpenAI.
Ampere CEO Renee James said in a statement today, “We started down this path six years ago because it is clear it is the right path.” Low power was previously associated with poor performance. That is untrue, as Ampere has shown. In an efficient computing envelope, we have pioneered the efficiency frontier of computing and provided performance beyond legacy CPUs.
