Baidu launched Anakin,Bitmain’s AI chip delivered strong performance
Baidu Create 2018 took place in Beijing on July 4th-5th, 2018. At the AI Computing System forum, which was a part of the event, Baidu launched Anakin v0.1.0, its latest AI inference acceleration engine. Anakin is highly flexible and able to run on multiple hardware platforms, including the SC3/SC3+ deep learning acceleration card which is based on Bitmain’s AI chip BM1682.
It is noteworthy that BM1682 outperformed all the other chips when testing Mobilenet in terms of latency. The chip’s performance was shown to be about 2 to 4 times higher than mainstream chips optimized by TensorRT or Anakin.
BM1682 comes with the customized hardware implementation for computing-intensive operator operations on convolutional neural networks and matrix multiplication neural networks, which explains why it boasts excellent performance. In addition, BM1682 has a large SRAM cache, which can achieve maximum graph fusion, significantly reduce data transmission and improve the utilization rate of computing units. That, in turn, puts BM1682 well ahead of most mainstream chips in testing.
BM1682 was launched in March 2018. With the 28nm process, its computing power is up to 3TFlops while the energy consumption is only 30W. As Bitmain’s second-generation AI chip for deep learning, it can work independently without an X86 CPU and support secondary development of customers. It is capable of 8-line decoding for H264/H265 and supports hardware acceleration for the post-processing of video images. Compared to the first generation, it consumes less energy and has higher density with fivefold higher performance. Moreover, it supports Ethernet and multi-chip interconnection over PCI-e for horizontal scalability.
As a high-performance computer chip maker, Bitmain regards the Internet industry as the major playground for its Sophon AI chips and products, which can be applied in Internet big data analysis, intelligent analysis of Internet videos, video advertising/marketing, and voice recognition/analysis, etc.
Wang Jun, Technical Director of Bitmain, said that, in the AI era, the Internet industry generates and gathers a large amount of data every day and it takes massive computing power to process such data with AI. In the future, Bitmain will keep it up and upgrade products at a fast pace to provide the industry with higher-performance chips and greater computing support.