Edge AI startup launches new-generation NPUs

September 17, 2018 // By Peter Clarke
Kneron Inc. (San Diego, CA) has announced a series of neural processing unit (NPU) intellectual property claiming it can achieve up to 13.17 tera operations per second per watt.

The KDP x20 series comprises three products the low power KDP 320, the standard KDP 520 and the high performance KDP 720. Power consumption is under 0.5W and performance is up o 5.8TOPS on a 28nm process at 600MHz clock frequency and with 8bit fixed-point resolution.

These are enhancements  to the KDP x00 series launched in March 2018 (see Qualcomm-backed startup announces AI processor family ).

These are Kneron's second generaton of cores and increase data movement efficiency, computation, memory utilization and has better optimized support for different neural networks. "The new NPU IP can be widely applied in diverse applications and satisfies complicated computing requirements," said Albert Liu, the founder and CEO of Kneron, in a statement.

The Kneron NPU IP allows neural networks to run on edge devices, including smartphones, home appliances, surveillance equipment, and a full range of IoT devices. The cores have an interleaving computation architecture plus compression technology that operates from the model layer down to the data and coefficient layers. Dynamic memory allocation enhances memory utilization without hampering computing performance.

The convolution layer can support both 8 and 16 bits fixed points concurrently to increase computing flexibility and the deep compression technology allows up to 50-fold compaction with less than a 1 percent impact on accuracy, Kneron claims.

www.kneron.com

Related articles:
Qualcomm-backed startup announces AI processor family
ST preps second neural network IC
Intel launches self-learning processor
ARM launches two machine learning processors


Vous êtes certain ?

Si vous désactivez les cookies, vous ne pouvez plus naviguer sur le site.

Vous allez être rediriger vers Google.