SamMobile has affiliate and sponsored partnerships. If you buy something through one of these links, we may earn a commission. Learn more.

Samsung develops new On-Device NPU tech for faster, smarter AI

General
By 

Last updated: July 2nd, 2019 at 16:27 UTC+02:00

As Samsung is pushing further with the development of its proprietary NPUs (Neural Processing Unit), the company has created a new high-speed, low-power On-Device AI deep learning algorithm promising 4 times more energy efficiency and 8 times more performance than existing solutions. The Samsung Advanced Institute of Technology (SAIT) detailed the new On-Device AI technology in a recent paper claiming that the algorithm has been successfully tested. It should pave the way for a future in which “all devices and sensor-based technologies are powered by AI.”

This increase in performance and energy efficiency has been achieved “by adjusting the data into groups of under 4 bits while maintaining accurate data recognition,” or what the company calls Quantization Interval Learning (QIL). Computations performed using the QIL process can achieve results similar to existing solutions while requiring 1/40 to 1/120 fewer transistors.

This method of compressing data into bit groups of 4 bits or lower allows for thousands of deep learning computations to be processed at the same time. It also allows for “and” and “or” computations to be performed in addition to arithmetic calculations such as multiplication.

Smartphones will become smarter and independent of the cloud with better On-Device NPUs

While Samsung hasn’t revealed exactly when this new solution will reach the market in consumer-grade devices, the company strongly hinted at its inclusion in smartphones at some point in the future. The new deep learning algorithm requires less energy and hardware, meaning that it can be adopted by On-Device NPUs “at the place where the data for an image or fingerprint sensor is being obtained.”

The Exynos 9820 chipset is the company’s first to feature an integrated proprietary NPU and it already expanded the capabilities of the Galaxy S10 series with enhanced photography and more advanced Augmented Reality features.

With a faster, less demanding On-Device solution now being in the works, we can only expect that features such as photography and AR will become smarter, faster, and more accurate with the new generation of devices.

But Samsung won’t stop at mobile chipsets. The company also revealed that it wants to apply this new deep learning algorithm to memory and sensor solutions in the near future. “We will live in a world where all devices and sensor-based technologies are powered by AI” according to the VP of Computer Vision Lab of SAIT, Chang-Kyu Choi. And the future is all about embracing On-Device AI technologies rather than relying too much on cloud-based AI solutions.

 

Source General Artificial IntelligenceNPU
Load 1 comments

You might also like

Samsung sets up an AI and data engineering lab for young millennials

Samsung, much like most other tech companies out there, is investing copious amounts of resources into machine …

  • By Anil Ganti
  • 11 months ago

Six new Samsung smart appliances certified for AI excellence

Half a dozen new smart home appliances from Samsung have just been certified for AI excellence. Over in the co …

  • By Dominik Bosnjak
  • 1 year ago

‘There will be no future’ if Samsung stops, chief says ahe …

As the world keeps suffering the effects of a historic pandemic, Samsung is embarking on a major hiring spree …

  • By Dominik Bosnjak
  • 2 years ago

Samsung to unveil a new AI product named NEON at CES 2020

Samsung will be unveiling a new AI product named NEON at CES 2020. The mysterious product is being developed b …

  • By Naresh
  • 2 years ago

Samsung launches a new web page for all its AI-related news

Samsung posts news about its products, collaborations, business, and more on its Samsung Newsroom portal. Thes …

  • By Naresh
  • 3 years ago

Samsung’s research shows realistic fake videos are not too far

In a paper titled Few-Shot Adversarial Learning of Realistic Neural Talking Head Models, researchers at the Sa …

  • By Naresh
  • 3 years ago