SAN JOSE, Calif. — A Beijing-based company that got its start designing ASICs for bitcoin mining announced it is sampling its first machine-learning accelerator. Bitmain Technologies’ BM1680 is optimized for both training and inference on deep neural networks.
The chip is sold in a fan-cooled module called the SC1. Bitman said it has been trained for Alexnet, Googlenet, VGG and Resnet neural networks and is compatible with Caffe, Darknet, Yolo and Yoto2 models as well.
Bitmain counts China’s big data center operators — Alibaba, Baidu and Tencent — among its top targets. It said it will also consider building its own machine-learning service.
The company plans to release technical details of the chip in a talk by its chief executive, Micree Zhan, at a Beijing event on November 8. It targets a wide range of use including image and speech recognition, autonomous vehicles and enhanced security cameras.
Bitmain was founded in 2013 and released its first chip, a SHA-256 accelerator for bitcoin mining, in November of that year. It claimed strong sales until late 2014, when the Mt. Gox exchange went bankrupt, tanking the bitcoin market.
The company made a comeback with a BM1384 chip as the bitcoin market rose in 2015. Executives claimed in a recent report they now have more than 600 employees, sell hundreds of thousands of mining systems a year and run their own mining business.
By the end of 2015, Bitmain decided to start work on a machine-learning chip. “Now after only a year and a half, we have the mass-production chips in hand,” said Zhan, in a press statement.
— Rick Merritt, Silicon Valley Bureau Chief, EE Times