Qualcomm Snapdragon Summit: Snapdragon 8 Gen 2 platform unlocks new AI capabilities

Qualcomm has officially unveiled the next iteration of its Snapdragon mobile CPU platform, the Snapdragon 8 Gen 2, promising a range of improvements with a particular focus on AI technologies.

Launched at the company’s annual Snapdragon Summit event in Hawaii, the Snapdragon 8 Gen 2 is ‘purpose built’ for AI, with an improved ‘AI Engine’ powered by an updated Qualcomm Hexagon processor.

Qualcomm said the chip has seen major architectural changes, such as the addition of a Tensor accelerator chip which is double the size of the previous generation, resulting in more than 4x faster AI performance and a 60% performance-per-watt improvement in power efficiency.

“We designed the [Snapdragon] 8 Gen 2 with one key technology: artificial intelligence,” said the company’s senior vice-president of engineering, Chris Patrick. “Qualcomm is the AI inferencing leader at the edge."

"Today we've already powered billions of edge devices across the world. We built on that expertise to make our new flagship an engine of innovation, a true AI marvel. We've integrated AI across the entire platform to usher in the new era of AI accelerated experiences.”

The new platform also includes an updated Qualcomm Kryo CPU with four performance cores and three efficiency cores, delivering up to speeds of 2.8GHz and 2GHz respectively.

A 3.2GHz prime core based on Arm’s Cortex-X3 technology also features, as does a new Adreno GPU with 25% improvements for both efficiency and speed. The company said this configuration offers up to 30% faster overall performance and up to 40% better power efficiency for longer battery life.

Another new feature introduced with this platform is the ‘Cognitive ISP’ - a physical link between the Qualcomm Spectra ISP and the Hexagon AI processor which uses neural network technology to automatically perform semantic segmentation.

This means that the camera itself can contextually recognise elements within a photo such as faces, hair, clothes and skies - and perform real-time adjustments to each one individually.

In addition to advanced on-device image processing, the AI functions supported by the new platform include multi-language translation, AI autofocus, and the world’s first support for INT4 on mobile, which essentially allows AI models to be trained on four-bit integers, rather than 32-bit floating-point integers.

According to Ziad Asghar, VP product management for mobile at Qualcomm, what this means is that more neural models will be able to run concurrently on the device.

“What we are providing is what we call Qualcomm AI Studio, which is a set of tools that allow you to take a model and actually shrink it down to integer four. And by doing that, now, instead of having to process 32 bits, you've got to process four bits only. That means you've got to do a lot less work, which means you've got to spend a lot less power.

“We are the first in the industry to do this,” Asghar said. “So it brings a huge advantage; a lot more AI is going to be able to fit in the same amount of processing, which is very, very powerful for us.”

Condensing 32-bit models to INT4 reduces the power consumption by a factor of 64, he said, and the company’s introduction of micro-tile inferencing also minimises the number of reads and writes for improved performance and longevity.

The company is aiming to attract more AI developers to the platform by unlocking these features, and is releasing a new developer kit to make it easier for them to start creating compatible machine learning applications.

The Qualcomm Innovators Development Kit will initially be made available to universities and research institutions before being rolled out on a wider basis in the future.

Qualcomm has also used these capabilities to add functionality to its Sensing Hub which uses dual AI processors to provide real-time contextual awareness by combining input from a range of sensors across the device, including cameras, microphones, gyroscopes, and more.

This will enable OEMs and developers to introduce a range of additional features, with some theoretical examples including the ability to identify the sound of glass breaking during the night to alert the user, and the ability to detect the orientation of the user’s face while lying down and prevent unwanted screen rotation.

Snapdragon 8 Gen 2 supports dual-active dual-SIM 5G, with a dedicated 5G AI processor to deliver better speed and stability throughput in both mmWave and sub-6GHz connections.

It’s also the first commercial chipset to support Wi-Fi 7 connectivity via the integrated FastConnect 7800 chip, taking advantage of the increasing number of Wi-Fi 7 routers coming to market. Qualcomm is boasting up to 5.8Gb/sec speeds, and less than 2ms latency, with cloud gaming and 4K streaming touted as key use cases.

Backing this up is support for the HDR10+ Adaptive standard, which compensates for changing lighting conditions when viewing HDR content, and OLED ageing compensation to keep OLED displays such as those found on the Samsung Galaxy S22 looking their best after prolonged use.

The Snapdragon 8 Gen 2 is also the first mobile chipset to support Unreal Engine 5’s Metahumans framework for creating and using hyper-realistic characters in games, and offers real-time hardware accelerated ray tracing, as well as 4K video playback at 60Hz.

Finally, Qualcomm’s AI tools are also being used to increase device security, including making facial unlock more secure by capturing data from facial unlock attempts, and routing it securely through to the digital signal processor (DSP) to perform detailed analysis on facial muscle movements and skin colour changes to ensure authenticity.

The new platform will be used in devices from major smartphone manufacturers including Oppo, OnePlus, Sony, and more, with the first Snapdragon 8 Gen 2 smartphones expected to hit shelves by the end of this year.