From MediaTek: "The future of smartphone AI is at the edge"
The era of practical artificial intelligence has truly arrived. In conjunction with the numerous sensors and cameras found in smartphones, edge AI processing has reached a point where AI-enhancements to applications are effective, immediate and - essentially - in demand!
What are the killer applications? Well, AI-enhancements can be applied to almost everything. Learning how a person uses their smartphone or app and reacting intelligently to tailor and improve the user experience, providing personalization to devices and make them more than just ‘out the box’. Neural network processing is now being readily used to augment and improve functions within apps such as video encoding, noise reduction and depth processing in photography, which MediaTek has already shown.
The challenges of bringing powerful edge AI processing into smartphones is not to be understated. Pure compute performance and memory bandwidth are design challenges, with considerations for thermal and power constraints of the device layered on top. These need to be met with innovative designs and software processing that allows a close-to-metal approach to avoid waste.
Edge AI processing offers essential advantages over purely Cloud-based processing such as always availability, immediacy, privacy and security. That's not to say remote processing is without merit, and it will continue to support edge processing because Cloud services offer oceans of deeper data to tap. However, the shift is clear: an ever greater need of on-device processing is the trend because applications are becoming increasingly more sophisticated, while usage scenarios layer ever more AI tasks, requiring AI processing resources to echo how a mutli-core CPU handles many threads. In fact, AI processing units like the APU 2.0 in the MediaTek Helio P90 already have a design akin to modern multi-core CPUs in that they use a hybrid mix of multiple core types that vary by degrees of performance and power-efficiency.
The APU 2.0 isn't the only AI task unit within the P90, though, the heterogeneous platform of CPU, GPU and APU gives app developers and device makers a unique flexibility that's also cross-platform so they can target any MediaTek chip, with or without an APU. What the P90 offers is an extreme performance based on a unique, in-house designed architecture that makes best use of available memory bandwidth and power. As such, the Helio P90 (12nm/32-bit LPDDR4X) out-performs industry flagship products (7nm/64-bit LPDDR4X) in widely used AI benchmarks.