Patented AI Chip Invades Deep Neural Networks

NovuMind has been awarded a US Patent for its NovuTensor artificial intelligence (AI) chip, a purpose-built processor for AI inferencing, with reported breakthrough performance-power ratios that make it viable for deployment in embedded applications and edge servers. US Patent No. 10,073,816, "Native Tensor Processor and Partitioning of Tensor Contractions," was issued on Sep. 11, 2018. 

 

NovuTensor is a special-purpose processor for AI inference applications. Designed for convolutional neural networks, it provides 15 tera-ops per second of computation, while consuming 5W. This makes it a suitable neural processing unit for embedded applications that require high performance and low latency, such as object detection for advanced driver assistance systems (ADAS). Reportedly, the processor outperforms other solutions because its design supports native 3D tensor computation on the chip. Less time is wasted moving data around and storing partial results, compared to other designs that rely on a series of 2-dimensional matrix operations.

Free Monthly Newsletter

Compelling read? Subscribe to FierceEmbeddedTech!

The embedded tech sector runs the market’s trends. FierceEmbeddedTech subscribers rely on our suite of newsletters as their must-read source for the latest news, developments and analysis impacting their world. Sign up today to get news and updates delivered to your inbox and read on the go.

 

NovuTensor is available as a chip or on a PCIe short card (174 mm) that can plug into a standard server. It comes with Linux-compatible software tools for development, runtime support and C++ API. For more info, visit NovuMind Inc.

Suggested Articles

Machine learning challenge will look for vocal communication between elephants and other behaviors

Iowa State University researchers are working with NSF grant

Brain Corp. reported a sharp increase in autonomous robot usage in 2Q