New edge AI chip Released by Bitmain for Cloud-Edge Integration
Beijing, October 17 - Bitmain today officially launched its Edge AI chip BM1880 with other products, including Sophon AI server SA3 and embedded mini AI machine SE3 which are based on Cloud AI chip BM1682, as well as EDB (Edge Computing Development Board), Edge AI module, and USB AI compute stick, which are all powered by BM1880.
Bitmain's Cloud AI chips fast iterate every 9 months. BM1682’s derivative products perform very well in actual operations, get positive feedbacks from customers, have extensive cooperation underway, and there are numerous security and surveillance projects being implemented according to Allen TANG, the Product Strategy Director at Bitmain. Meanwhile, in order to better meet the customer and market’s demand, Bitmain is developing and rolling out AI chips and solutions at the edge for the target markets requiring "Cloud-Edge" integrated AI solutions.
From Cloud to Edge, chip products fast iterate every 9 months
The BM1682 chip was launched in March 2018 The product offers 3 TFLOPS peak performance and 30Wpower consumption, and it is the second-generation AI chip for deep learning released by Bitmain. It can operate independently without X86 CPU, providing support for customers’ development, and offering 8-channel H.264 and H.265 decoding capabilities on a single chip. It also enables hardware acceleration for video/image post-processing, and provides support for multichip interconnection via Ethernet and PCIE for the scalability. Compared to the first generation, the BM1682 chip features a higher density and significantly increases the actual performance.
Based on BM1682 chip, Bitmain has developed a variety of products, including the deep learning accelerator card SC3, AI server SA3, facial recognition server SS3-C3, and embedded mini AI machine SE3, which are widely used in security and surveillance, internet, smart parks and other areas.
The BM1880 chip, which was taped out successfully in July 2018, is an AI chip for deep learning inference focusing on edge applications. It delivers a peak performance of 1 TOPS for 8-bit integer operations and up to 2 TOPS with Winograd convolution acceleration. In addition, the specially designed TPU scheduling engine supplies extremely high bandwidth data flow effectively to all the tensor processor cores. This chip is equipped with 2 MB on-chip memory to offer the best programming flexibility for performance optimization and data reuse. This product is also provided with a powerful deep learning model compilers and SDK. Mainstream deep learning frameworks, such as Caffe and Tensorflow can be easily ported to the BM1880 platform. Popular neural network models, including CNN, RNN and DNN can be run on this platform.
BM1880 can be used as a co-processor for accelerating deep learninginference. In addition, it can be used as a main processor to receive video streams, images, and other data from Ethernet or USB interface and carry out inference and other computer vision tasks. This product can also receive video streams and image data from other hosts and then carry out inference and return the results back to the hosts.
At the media briefing, TANG Weiwei also revealed the development of the next-generation Cloud chip BM1684, which is expected to be launched with significantly increased performance at the end of the year. The new server SA5 based on BM1684 will also enable a greatly increased number of video structuration channels, far exceeding similar products.
High-performance AI Server SA3 with focus on AI and security & surveillance
Bitmain Sophon AI Server SA3 is developed based on the second-generation AI chip BM1682. In this series, the SA3-23 server is a 19" 2U high density server, equipped with 3 intelligent processing units, each of which is comprised of 6 BM1682 chips. SA3 has the following advantages:
- High Performance
SA3 includes an interconnected cluster architecture incorporating a number of BM1682 chips, offers low-power main control technology, and delivers cluster management, task and job scheduling, load balancing, video and image API interfaces and other services, with parallel computation ability for massive video data, stable, reliable and timely response to external requests, and manageability.
- High Energy Efficiency
Compared to those servers designed by traditional vendors relying on general-purpose GPUs, the SA3 server based on BM1682 offers more than 2 times Performance per Watt , while meeting the needs for high-performance video analysis.
- High Computation Density
AI Server SA3 also greatly reduces the required computation space, while delivering super computing power with lower overall power consumption. In comparison to other mainstream 4U servers with comparable computing power in the industry, the 2U height of SA3 needs less rack space in actual machine room deployment to allow the customer to select a more flexible deployment scheme, so as to reduce the total cost of ownership (TCO) for data center deployment, achieving the goal of lowering the operation costs and empowering the industry at the same time.
It is worth noting that SA3 video structuration server (2U) enables up to 90 video structuration channels, delivering 2.2 times more performance than a traditional industry server (based on GPU), and three times more performance for facialrecognition than a traditional server (based on GPU).
Sophon AI Server SA3
In addition, Bitmain is actively exploring the smart park market in order to enable the smart park application scenarios with AI. The embedded mini AI machine SE3 based on BM1682 offers a compact size of 210mm*115mm*45mm, a single-precision peak performance of 3 TFLOPS, and a rated power consumption of 60W. It provides support for 4 dynamic or 10 static facial recognition streams, far exceeding other similar products in the market, thus gaining extensive customer’s attention.
Embedded Mini AI Machine SE3
AI terminals launched for edge computing
In addition to cloud AI chips, Bitmain is also aggressively expanding its new product lines. A number of AI terminals which are based on the Edge AI chip BM1880 were launched simultaneously.
Sophon EDBis specifically designed for all kinds of edge computing applications that require powerful deep learning capability. With itscomplete prototype designs, it facilitates developers in carrying out various types of application development. This product is equipped with one BM1880 chip, and provides support for Arduino and Raspberry Pi development modules. Developers can utilize the these modules to develop deep learning and machine vision-related applications, including face detection and recognition, facial expression analysis, object detection and recognition, license plate recognition, and voiceprint recognition.
Edge Computing Development Board
Bitmain further designed two products, i.e., AI module and AI compute stick, which are used for various edge applications through USB interface. Both products incorporate the BM1880 chip and enable AI functions on conventional products through USB interface. They can be widely used on smart IP camera, robots, industrial computers, and etc.
USB-Connected AI Compute Stick