Mipsology, an innovative neural network acceleration startup, today announced its Zebra neural network accelerator running on Advantech’s VEGA-4001 acceleration boards achieved the highest throughput for a neural network-based image classification task.

The integrated solution, using two Xilinx VU9P FPGAs, 32GB DDR memory and a 16x PCIe interface, can process 25,000 images per second using CaffeNet under TensorFlow, Caffe, Caffe2 or MXNet. Zebra can process most convoluted neural networks, such as ResNet, Inception or VGG, without any change, and still deliver high throughput on Advantech VEGA-4001.

“These results thrust VEGA and Zebra to the leading position in inference performance,” says Ludovic Larzul, founder and chief executing officer of Mipsology. “The combination of VEGA-4001 boards’ high throughput suitable for data center applications and Zebra’s performance and ease of use will replace GPUs and CPUs for inference without changing the neural network, the framework or the application.”

“We are impressed by Mipsology’s engineering talent and ability to use FPGAs to accelerate neural networks,” notes Brian Carr, product strategist and marketing director for Advantech Video Solutions Division. “Our VEGA-4001 boards combined with Zebra are a compelling solution for accelerating the heavy computation of deep learning inference in data centers and a perfect power efficient replacement for large GPUs.”

The VEGA-4001 board is sampling now and will be generally available in first quarter of 2019. More information can be found at www.advantech.com

Advantech and Mipsology will exhibit at SC18 Monday, November 12, through Thursday, November 15. Advantech will be in Booth #32018. Mipsology will demonstrate Zebra results in Booth #3871 as part of the Startup Pavilion.

About Mipsology

Mipsology is a startup developing state-of-the-art FPGA-based accelerators targeted for deep learning applications in neural networks. It was founded in 2015 by a team of engineers and scientists who created a family of world-class FPGA-based super-computers over the past 20 years. More information is available at www.mipsology.com