Arm Ltd., the British firm whose semiconductor designs are used in most of the world’s connected devices and handsets, today debuted the chips that will likely power the next generation of flagship Android phones.

The unveiling comes hot on the heels of the company introducing a new accelerator for virtual reality headsets. While that chip, known as the Mali-D77, focuses mainly on improving graphics quality, the common theme in the latest additions to Arm’s portfolio is that they all promise to boost artificial intelligence performance dramatically, in some cases in edge devices, a growing trend. That’s accompanied by several major enhancements in other areas.

Cortex-A77

The Cortex-A77 central processing unit is debuting less than a year after its predecessor, the Cortex-A76, which powers Samsung Electronics Co. Ltd.’s flagship Galaxy S10 phones. Designs from the mobile CPU series are also used in a range of other top-end devices including the Google LLC-made Pixel 3 series.

The Cortex-A77 features upgrades that enable it to provide 20% more single-thread processing power than the previous design. Moreover, the chip is 35% faster when performing floating-point operations, the calculations that AI models use to process data. Arm claims that the result is up to 35 times better machine learning performance than a Cortex design from two years ago.

In addition to apps that incorporate AI models, the company sees the CPU’s increased speed benefiting other resource-intensive mobile workloads such as augmented reality services.

“ML use cases on mobile devices are becoming more complex, so having a CPU that can support this increasing compute demand is vital,” explained Stefan Rosinger, Arm’s product management director for the Cortex-A series. “Just some of these use cases on devices include AI cameras, visual scene detection, 3D scanning, biometric user ID (face recognition), voice recognition, ML in gaming and ML in AR.”

Analyst Holger Mueller of Constellation Research Inc. told SiliconANGLE that although the Cortex-A77 looks promising, it’s still too early to tell what the best mobile chip for 2019/2020 will be. However, he said that Arm’s capabilities in power consumption and the ability to deliver AI capabilities on a chip are good true Norths in terms of mobile chip development.

“We’ll have to wait to see the next handset designs [based on the Cortex-A77] in a few months to get a better idea of how successful Arm will be, but certainly it wants to have a bigger piece of the mobile chip market,” Mueller said.

New mobile GPU

Arm today also introduced the Mali-G77 graphics processing unit, which boasts even more significant enhancements than the new CPU. The company expects the chip to deliver no less than 40% higher peak performance than the previous-generation Mali-676, the GPU included in the Galaxy S10 family.

Arm said that the Mali-G77 provides this speed boost while consuming 30% less power. AI apps, in turn, are expected to run up to 60% faster than on the Mali-676, which should enable developers to implement more machine learning features locally on the user’s device.

Andy Craigen, the director of product manager for the Mali-G77, explained that “being able to perform tasks on the device rather than sending them to the cloud for processing means improved performance, less latency and fewer security concerns.”

Another key improvement is that the Mali-G77 boasts 30% better performance density. This means GPUs based on the design will take up that much less space, which is significant because the graphics card is usually one of the largest components on a phone’s chipset. The extra room will enable device makers to pack more silicon into their handsets and thus further improve performance.

The Mali-G77’s impressive specifications are the result of drastically redesigned internals. Arm has based the GPU on a brand new architecture called Valhall and upgraded several core components, not just the ones focused on AI. One of the most significant enhancements has been made to the texture mapper, a module that plays a key role in rendering complex visual assets such as mobile game graphics. 

“The improved gaming performance of Mali-G77 is linked to the quad texture mapper, which provides four texels/[processor] cycle. This is 2x greater throughput than Mali-G76 and 4x greater than Mali-G72,” Arm’s Andy Craigen wrote. “The quad texture mapper provides improvements across the board of high-fidelity and casual gaming, but has an especially large impact on texture-heavy games.”

Second-gen ML accelerator

Rounding out the trio of products unveiled today is a new iteration of the Arm Machine Learning Processor. It’s a chip architecture born from Project Trillium that is built for the sole purpose of running AI workloads, much like Google LLC’s Tensor Processing Unit. The difference is that Arm is targeting a much broader range of systems, from handsets to industrial robots and data center servers.

The second-generation Machine Learning Processor represents a notable milestone for the company’s efforts in this area. Arm has made the design much more scalable, allowing manufacturing  partners to include up to 8 processing core in a chip for a maximum speed of 32 trillion operations per second. That’s up from a peak performance of 4.6 trillion operations per second before.

Power utilization is better as well, with Arm claiming that it has more than doubled the architecture’s energy efficiency. The company also promises a threefold improvement in memory compression, a data management technique that allows a chip to process more information at once by shaving off disposable bits.

Arm has already released the new designs to manufacturing partners and expects the first chips to start shipping in devices next year.

Photo: Arm