// php echo do_shortcode (‘[responsivevoice_button voice=”US English Male” buttontext=”Listen to Post”]’)?>
Edge AI chip startup Deep Vision has raised $ 35 million in a Series B financing round led by Tiger Global and existing investors Exfinity Venture Partners, Silicon Motion and Western Digital.
The company began shipping its first generation chip last year. ARA-1 is designed for energy-efficient, low-latency edge AI processing in applications such as smart retail, smart city, and robotics. While the company’s name suggests a focus on convolutional neural networks, ARA-1 can also accelerate natural language processing with support for complex networks such as long short-term memory (LSTMs) and recurrent neural networks (RNNs).
A second generation chip, ARA-2, with additional features to accelerate LSTMs and RNNs, will hit the market next year.
“When ARA-1 started sampling last year, we wanted to make sure we had a fantastic product-market fit with customers,” said Ravi Annavajjhala, CEO of Deep Vision EE times. âWe have several customers including a large FAANG customer that we ship to in bulk, our product is qualified, it is in mass production, and now we are taking this story and replicating it across different segments and different customers. â
“FAANG” refers to the hyperscalers Facebook, Amazon, Apple, Netflix and Google.
Annavajjhala said smart retail is the company’s main target application, adding that the company already has a “very large customer” in the sector along with “several other commitments”. He declined to identify the retail customer or whether he and the FAANG company were one and the same.
Retail applications for Deep Vision’s chips include cash register analysis and inventory management, including shelf cameras and digital signage. While shelf cams don’t need extremely low latency, according to Annavajjhala, they need millions of images to track products via fairly complex AI models. Therefore, the processing requirements are relatively high.
âA shelf cam is not only [detecting empty shelves]”It combines this model with a number of other models, often with many filters,” he said. âComputing itself is a problem – we can’t bring it to the cloud because it’s prohibitively expensive – but the network bandwidth and power over Ethernet switches required have also become very expensive.â So, âit’s in the best interests of the business to reduce the total cost of ownership, as far as possible on the verge of conclusion. “
Other applications are Smart City (surveillance cameras with high resolution and high frame rate), driver monitoring systems, robotics, drones and factory automation.
ARA-1’s performance for ResNet-50 was 100 frames per second or 40 frames per second per watt. The chip is now supplied in USB modules, M.2 modules and U.2 PCIe modules (two or four chips on one card). The company offers two SKUs of the chip – an 800 MHz version and a 600 MHz version for performance-sensitive edge applications.
Deep Vision’s computing architecture is designed to minimize the amount of data movement to and from storage.
“The data movement is minimized across software, at the system level and at the computing kernel level to ensure that all data transferred to the chip remains as close as possible to the computing kernel and is retained for as long as possible,” said Annavajjhala. “We abstracted all of this on the hardware and made sure that the data flows minimize the movement of data between compute and different levels of the storage hierarchy.”
Data can be reused many times in computing AI inference, he said, noting that with image convolutions, about 90 percent of the data is the same between one frame (or convolution) and the next. Deep Vision’s software scans models and looks at many different combinations of data movement and plans the most efficient combination for the processor. A compiler that tracks the power consumption and the performance of the processing and storage subunits, optimizes for the desired result. Deep Vision’s software flow supports TensorFlow, Kaffe, Pytorch, MXNet and ONNX.
The time pressure in bringing ARA-1 to market resulted in some advanced features being left out, according to Annavajjhala. These functions, mainly related to the acceleration of LSTMs and RNNs, are added in the ARA-2 chip, which is based on the same overall architecture. ARA-2 will also switch from 28 to 16 nm process technology. The result will be a 3-fold increase in performance per watt, which is a 5-fold increase in performance compared to ARA-1. (Applications currently running on ARA-1 will be compatible with ARA-2, the company said).
Deep Vision was founded in 2018 by Doctors Rehan Hameed and Wajahat Qadeer from Stanford University and has now raised a total of 54 million US dollars. The company currently employs 57 people and is expected to grow to 75 by the end of the year. ARA-1 is shipping in bulk today, while ARA-2 is expected to begin sampling in 2022.