Machine Learning Start-up Taps Alpha-Data's FPGA Technology to Reach 750 AlexNet-like Inferences per second on Single Low-Profile Xilinx FPGA boardSANTA CLARA, Calif., Oct. 25, 2016 — (PRNewswire) — TeraDeep, a developer of accelerated deep learning applications and AI appliances, announced another significant performance milestone for its TeraDeep Accel™ FPGA technology used for video analytics in on-premise appliances for security and protection of high-value assets.
The TeraDeep Accel technology for single-FPGA applications is currently used for low-power, low-latency on-premise appliances. The company completed development of a traditional AlexNet-based solution that can run low-latency applications with up to 750 inferences per second (IFS) on a Kintex Ultrascale 115 FPGA board. This represents close to a 2X performance advantage compared with large GPUs, while consuming 5 times less power. Moreover, compared with an Intel Xeon E5, the current TeraDeep Accel technology delivers 10 times the performance while consuming 5 times less power.
The Alpha-Data board, which runs TeraDeep's deep learning acceleration technology, provides another single-FPGA option in a small form factor for the company's upcoming AI appliance. AI appliances are expected to help bring about the development of "AI Cities," an extension of "Smart Cities."
"We found the low-profile Alpha-Data boards enable us to deliver great performance in a small form factor for low-key on-premises appliances used in security and safety applications," said Didier Lacroix, TeraDeep's CEO. "We are running elaborate Deep Learning detection algorithms that would otherwise require the use of traditional GPUs that are 4 times larger and 10 times more power hungry."
Accelerating Deep Learning Development for FPGA-based Appliances and Servers
The results were achieved through TeraDeep's suite of advanced deep learning neural network algorithms, optimized to run on the company's TD Accel™ acceleration platform.
TeraDeep's collaboration with Alpha-Data extended to being early adopters of Alpha-Data's new KU115-based boards. The large amount of on-board compute capabilities enabled TeraDeep to explore the scaling of its TD Accel technology.
"High-capacity FPGA boards are key to increasing the compute density of Deep Learning solutions," said David Miller, Managing Director at Alpha-Data. "TeraDeep's use of Alpha-Data technology demonstrates its capabilities for demanding, high-performance Deep Learning applications."
TeraDeep is now engaged with several partners on the development of next-generation heterogeneous architectures that will deliver other significant performance improvements for appliances and emerging data center applications.
TeraDeep literature will be available at Alpha-Data's booth #2607 at the Supercomputing 2016 conference November 14-17 in Salt Lake City, Utah.
TeraDeep is a pioneer in the development of specialized neural networks and acceleration technology for deep learning. The company develops and markets unique AI Appliances offering complete solutions embodying a creative combination of specialized hardware designed for purpose-trained networks, available through a robust set of APIs. TeraDeep AI Appliances are leased either directly or through its value-added partners. It licenses reference designs and pursues OEM models with qualified parties.
To view the original version on PR Newswire, visit: http://www.prnewswire.com/news-releases/teradeeps-fpga-based-ai-inference-board-speeds-hd-video-analytics-for-on-premise-security-appliances-300350492.html
For more information contact: Tom Breunig, Helios Communications, (503) 858-2801