Like many of the emerging neural network processor chips the NovuTensor is designed for inference computations to be performed at the edge of the network.
The NovuTensor architecture is based on a patented design that can perform three-dimensional tensor computations directly, avoiding the overhead inherent in other chips which require tensors to be unfolded into 2-dimensional matrices. This approach, covered by US Patent No. 10,073,816, is optimized for convolution-based deep neural networks, the company said.
NovuMind has performed benchmark tests on a variety of networks, including ResNet18, ResNet34, ResNet50, ResNet70, VGG16, and Yolo2. And compared them against leading GPUs used for NNs.
Ren Wu, founder and CEO of NovuMind, made the point in a statement that while GPUs perform reasonably well when processing large batches of data, these chips are not suited for real-time applications, single inferences and small batch sizes and low latency. "They also lack power efficiency and they tend to be very expensive. With the arrival of our NovuTensor chip, we are breaking these barriers and ushering in a new era where AI can be deployed at scale," he said.
NovuMind has produced the benchmark data from first tests on the 28nm silicon.
"We produced our first chip using a conservative 28nm semiconductor process, to validate the design. I am pleased to announce that the first chips work perfectly and our testing validates the superior performance we were expecting. We outperform the most advanced chips from the competition, even those costing thousands of dollars. As we migrate our design to more advanced semiconductor processes such as 16nm or 7nm, our advantage will extend even further," Wu said.
NovuMind was founded in 2015 by Wu, formerly a distinguished scientist at Baidu, with 50 people, including 35 engineers working in the U.S. and 15 in Beijing.