The A78 is the company's most powerful IC yet for smartphones, but Arm is finally taking orders for even more muscular cores...
Arm announced its next top-of-the-line CPU for smartphones – the A78, along with its latest Mali graphics processing unit (GPU) and Ethos neural processing unit (NPU). It also revealed a program in which the company is customizing its leading-edge cores for customers.
Arm’s new Cortex-A78 delivers a 20% increase in sustained performance over its predecessor, the A77. That should be no surprise; the company built its reputation on power efficiency, and remains focused on that metric. Part of the performance gain comes from jumping down a node. The A77 was built in a 7nm process; the A78 at 5nm.
Delicate tuning is also required to achieve that kind of a boost. Arm says the A78 has greater on-device machine learning (ML) performance, and manages compute workloads more efficiently.
It will actually throttle back the frequency of certain subsystems to stay within the commonly desired power envelope of 1 watt., explained Ian Smythe, vice president of solutions marketing in Arm’s Client Line of Business. The processor will find a way to perform that tradeoff in a manner that will not compromise the performance of apps, he said.
Even so, a tradeoff is a tradeoff, and some Arm customers prefer to have a processor that is capable of running as fast and as hard as possible —one that can occasionally kick into Beast Mode. And if that plays hob with the thermal budget, they’ll accept the consequences.
That’s what the Cortex-X program is for. Arm will customize its Cortex cores to meet performance targets specified by customers for flagship smartphones and large-screen devices. Arm has always done some customization for customers before, but this is different, Smythe said.
“This is outside the roadmap,” he noted.
He said that several customers contributed to the specification of the first Cortex-X core – the Cortex-X1. The Cortex-X1 will have a 30% peak performance increase over Cortex-A77.
Smythe declined to identify who the customers were. They are no doubt smartphone manufacturers hoping to close the performance gap with Apple, which designs its phone processors itself and remains the performance leader in that category. Arm’s announcement was accompanied by an endorsement of the Cortex-X program from Samsung – take that for what it might be worth.
Smythe said the company has multiple partners that are using A78. He said Arm is planning to deliver silicon this year. Arm’s customers’ schedules are their own, he said, but he anticipates the first devices built around the A78 are likely to start appearing early next year. He said the production of the Cortex-X1 and the appearance of the first products integrating it will be “in same time frame.”
Mali and Ethos
At the same time, the company introduced the Mali-G78 GPU, which Arm described as its highest-performing Mali GPU based on Valhall architecture, with a 25% increase in performance. G78 GPUs can support up to 24 cores.
The company attributed the performance gain to asynchronous, top-level tiler enhancements, and improved fragment dependency tracking.
“And based on demand from partners, alongside the premium Mali-G78, we are introducing a new sub-premium tier of GPUs. The first in this new tier is the Arm Mali-G68, which supports up to 6 cores and inherits all the latest Mali-G78 features, enabling our partners to deliver high-performance experiences to a wider audience of developers and consumers,” the company said.
The new Arm Ethos-N78 NPU boasts a 25% improvement in performance efficiency from its predecessor. The company said it was designed to address ML use cases ranging from AR-based smartphone applications to smart home-hubs. Arm is introducing the Ethos-N78 neural processing unit (NPU). Configurations are flexible, and range from 1 tera operations per second to 10 TOPS. Smythe said the N78 provides better than 2x peak performance; the figure of merit there, he said, is inferences per second per square millimeter.
[Note: this article has been edited. We originally misspelled Ian Smythe’s name minus the terminal E. We also identified the new custom design as the Cortex-1; it is the Cortex-X1. -ed.]