The Snapdragon 888 platform, unveiled on December 2 at Snapdragon Tech Summit Digital 2020, features a revised architecture, 5G multi-SIM connectivity, 5nm processes, a Kryo 680 processor (up to 2.84GHz), a third-generation Snapdragon X60 5G Modem-RF system, 6th-gen Qualcomm AI Engine, and a Triple Image Signal Processor (ISP) -- as well as Spectra 580 -- to data capture at up to 2.7 gigapixels per second.
The latest flagship processor is expected to feature in premium mobile devices in 2021. While the revamped image processing functionality could be a way for OEMs to separate their camera-equipped smartphones from other models in a congested market, the AI functionality of the new processor family makes it possible.
Jeff Gehlhaar, VP of Technology, and Hsin-I Hsu, Senior Product Manager at Qualcomm said the 6th-generation Qualcomm AI Engine is the new jewel in Qualcomm's mobile AI crown, made possible through a substantial redesign of the AI processing system.
As artificial intelligence becomes more complex -- now featured in our camera software, games, the natural language processing (NLP) capabilities of apps, and more -- an engine capable of high-performance AI processing is necessary to prompt innovation in the space.
Qualcomm's Hexagon 780 processor is central to the redesign of the engine. The Hexagon 780 has been dubbed a "fused AI accelerator [which] features our biggest leap in architecture and performance in years" by the US chip giant, with prominent changes in how the system operates resulting in improved AI processing.
Qualcomm used scalar, vector, and tensor accelerators in its Hexagon chip. The scalar accelerator is now 50% more powerful, the tensor accelerator is twice as fast as its Snapdragon 865 predecessor, and the Qualcomm Hexagon Vector eXtensions (HVX) has been improved in order to support additional data types.
The company has also removed the physical distance between each accelerator by 'fusing' them together, creating one larger AI accelerator unit. Each accelerator now shares a dedicated memory bank to share and transfer information more efficiently.
Qualcomm says the memory bank is 16 times larger than previous designs, and this has resulted in hands-off times being up to 1,000 times faster in "specific" use cases.
In addition, Qualcomm has bumped up the Adreno 680 GPU with new instruction sets for rapid calculations, and in total, improvements of 43% in performance have been reported during tests.
The Snapdragon 888 utilizes the 2nd-generation Qualcomm Sensing Hub, which now includes an always-on, low-power dedicated AI processor. This additional hardware allows up to 80% of a workload to be offloaded from the Hexagon processor, improving power consumption rates, Qualcomm says, and also contributes to an overall AI improvement in Snapdragon 888 of five times in comparison to previous designs.
TOPS is the measurement to work out how many computations an AI chip can handle at the same time. While the maximum number of operations is only part of the story, a TOPS rating is considered a benchmark in the industry. According to Qualcomm, the Snapdragon 865 achieved 15 TOPS, whereas the Snapdragon 888 is capable of 26 TOPS.
Qualcomm is also working with Google to optimize the TensorFlow machine learning framework on the new chip in the sensing hub.
The "significant upgrade" to Qualcomm's AI engine has been backed by improvements to the firm's Neural Processing SDK, which has been revamped to support additional models and Microsoft Windows 10 AI use cases. Qualcomm is also collaborating with Hugging Face, a natural language processing developer, to improve the firm's NLP library.
The 6th-gen AI Engine's API, which now gives developers access to the Hexagon 780 processor, Adreno GPU, and Kryo CPU, has also been made backward compatible with the 5th-gen Qualcomm AI Engine.
"We're introducing the Qualcomm AI Engine direct with our 6th-gen Qualcomm AI Engine, moving to the next chapter of our AI software," Gehlhaar and Hsu said. "With this solution, we extend and enhance the capabilities of our AI software solutions to provide developers with access directly to the hardware, and not only for the Hexagon 780 processor but also for the Adreno GPU and Qualcomm Kryo CPU."