Edit | Vincent
“Today, Horizon Robotics held a product launch event where it unveiled its self-developed embedded intelligent vision chip and three intelligent solutions for intelligent driving, intelligent city and intelligent business.
This is the first time horizon has publicly announced a new product after a long hiatus, and for those who have been following Horizon for a long time, this is no doubt a bombshell. After leaving Baidu, Kai Yu, who has embarked on the road of entrepreneurship, finally demonstrated his determination and ambition of Horizon layout AI with his strength today. Can he, who has been looking at horizon for a long time, see his own brilliance as he wishes?”
14:00 PM, December 20, 2017, China Grand Hotel, Beijing.
Horizon Robotics released what it called China’s first and the world’s leading embedded ARTIFICIAL intelligence vision CHIP BPU (Brain Processing Unit) at its product release conference: Sunrise processor for intelligent camera and Journey processor for intelligent driving.
Dr. Kai Yu, founder of Horizon Robotics technology and former leader of Baidu Deep Learning Research Institute, said at the meeting that Horizon took the lead in the industry to propose to build ARTIFICIAL intelligence chip architecture (BPU) as early as 2015 when it was founded, which was earlier than Google’s TPU architecture.
According to the introduction, the chip power consumption can be as low as 1.5W, can simultaneously pedestrians, motor vehicles, non-motor vehicles, lane lines, traffic signs, traffic lights and other eight types of targets for accurate real-time detection and recognition.
In addition, Horizon line also unveiled three smart solutions at the conference, which are aimed at: Intelligent driving, intelligent city and intelligent business, the latter two solutions will apply horizon’s embedded artificial intelligence vision chip, set unique deep learning algorithm, in the front end can achieve large-scale face detection and tracking, video structure and other applications, can be widely used in intelligent security, smart city and other scenes.
AI Front was on the scene to get detailed specs for Horizon’s chip launch:
Based on Gauss architecture, the chip can complete a video input of 1080P@30fps and can detect and identify 200 targets in each frame of image. Typical power consumption is only 1.5W, and the delay is less than 30ms.
At the same time, Dr. Kai Yu also announced Horizon’s strategy in the chip field in the next few years:
In 2018, Horizon will launch a new smart chip based on Bernoulli architecture with a new sparse binary neural network structure. In 2019, a high-performance chip capable of simultaneously processing 12 channels of 4K@30fps video input will be launched.
By 2025, each of the 30 million new cars on China’s roads will have autonomous driving capabilities, each with a “brain” based on Horizon’s ARTIFICIAL intelligence processor.
Horizon’s long-planned “big move”
Since its establishment in July 2015, Horizon has released two intelligent systems, lane/vehicle/pedestrian detection intelligent driver assistance system and Andersen Platform, which are respectively applied in intelligent driving and smart home. In July 2016, after obtaining a new round of financing, Horizon said that the investment would be used to increase research and development investment in autonomous driving and smart home, and speed up product development and delivery; To promote the research and development of artificial intelligence chips and systems. In September 2016, Horizon set up the Nanjing R&D center, and the development of smart chips began to get on the right track.
For more than a year from September 2016 to December 2017, Horizon seemed to be missing from the field of technology, and no information was released to the outside world. It even caused the outside world to wonder whether Horizon could continue to survive in the field of AI, which is now like a battlefield.
Finally, in December 2017, AI Front received an invitation from Horizon to attend this product launch event. When we learned that the new product is most likely to be a chip, we realized that horizon, which has been working hard and silent, is brewing such a “big move”.
Chip big move has been issued, can the battlefield kill?
Now, horizon’s chip has also been released, the big move has been issued, but in this has blown the storm of intelligent chip battlefield, can play a kill effect? Throughout the world, technology enterprises in the field of smart chips have occupied their respective mountains. Let’s take a look at the current situation of the AI battlefield:
Foreign battlefields are mainly dominated by dachang. Intel, Nvidia, Google and Microsoft, four old technology giants, occupy fertile ground respectively, layout AI and grow crazily, quite a situation of hegemony among giants.
Intel – An old chip maker trying to reinvent itself
Autonomous vehicles have become one of the main targets of artificial intelligence, and Intel wants to solidify its position in the field.
However, rather than focusing exclusively on internal research and development, Intel is building its AI capabilities through acquisitions. In August 2016, Intel acquired Nervana Systems, a maker of neural network processors.
In November 2016, a few months after the acquisition of Nervana, Intel announced the launch of a series of processors, Nervana, a platform aimed directly at AI related applications such as training neural networks. “We expect Intel’s Nervana platform to deliver performance breakthroughs and significantly reduce the time required to train complex neural networks,” said Diane Bryant, Executive vice president and general manager of Intel’s Data Center group. Intel’s performance is expected to improve 100-fold within a decade, accelerating the pace of innovation in the emerging field of deep learning. “
In March 2017, Intel made a high-profile acquisition of Deep learning ADAS developer Mobileye for about $15 billion. Intel’s acquisition strategy was almost immediately significant. The chipmaker wants to gain a foothold in autonomous vehicles, a strategy that has cataputed it into the role of a key supplier of machine learning hardware.
At The Automobility LA trade show in Los Angeles in November, Intel’s CEO Brian Krzanich announced that autonomous driving has become the biggest game changer of the day. It announced that Intel’s new products, SoC and EyeQ5, following its acquisition of Mobileye, are more than twice as powerful as its closest competitor, Nvidia’s Xavier’s deep learning platform.
Nvidia — Dominates the GPU market
While there are plenty of GPU companies in the market, none is more synonymous with the technology than Nvidia. According to Jon Peddie Research, Nvidia’s GPU shipments grew 29.53 percent in the third quarter of 2017, as archrivals AMD and Intel both lost ground. AMD’s shipments increased 7.63 percent, while Intel’s shipments increased 5.01 percent. Of course, this is largely driven by the video game market, but analysts at Jon Peddie Research believe demand for high-end performance from cryptocurrency-mining related apps is also contributing to the shipment growth.
The need for high-performance task processors, such as cryptocurrency mining and AI applications, has pushed Gpus to the forefront of AI hardware. Gpus contain hundreds of cores that can execute thousands of software threads simultaneously and are more energy efficient than cpus. CPU is more general, and more jump, can perform many tasks, and is good at repeated operations on large amounts of data. Gpus are called gpus because of this key distinction and are better at handling graphics — as graphics involves processing thousands of small calculations at once. At the same time, such performance also makes gpus ideal for understanding tasks such as neural network training described above.
Just back in December, Nvidia announced the release of Titan V, a PC GPU designed for deep learning. Based on Nvidia’s Volta architecture, the GPU uses a new core technology Nvidia calls Tensor Cores. In mathematical terms, a tensor is defined as “a mathematical object similar to a vector, but more general, represented by a set of spatial coordinate functions.” What Nvidia has done is develop cores with complex architectures to handle the needs of deep learning and neural network computing.
Google — focus on TPU and do a good job in application implementation
Perhaps no company has studied the concept of tensors more deeply than Google. In 2016, the search giant released TensorFlow, a very popular open-source framework for deep learning. As Google says, “TensorFlow is an open source software library that uses data flow diagrams for numerical calculations. The nodes in the diagram represent mathematical operations, and the edges of the image represent multidimensional arrays (tensors) that communicate between them. Its flexible architecture lets users deploy computing to one or more cpus or Gpus on a desktop, server, or mobile device using a single API. “
In 2016, the company released the first generation of a new processor called a tensor processing unit (TPU). Google’s TPU is an ASIC tailored specifically for machine learning and TensorFlow. In May 2017, Google released the second-generation TPU and said it had 180 teraflops performance.
In June 2017, at the 44th International Symposium on Computer Architecture (ISCA) in Toronto, Canada, Google presented a study that combined its TPUS with Intel Haswell cpus deployed in data centers, And Nvidia K80 Gpus deployed in the same data center. Tpus were found to run on average 15 to 30 times faster than Gpus and cpus. The TOPS per watt of TPU is also 30 to 80 times higher than the latter two. TPU is now available in all of Google’s online services, such as search, Street View, Google Albums and Google Translate, Google said.
TPU has also proven its value in some very high-end AI applications. TPU is the “brain” behind Google’s famed AlphaGo AI, which beat the go world champion last year and recently made a giant leap forward in artificial intelligence by proving it can become a go master on its own in a relatively short period of time. After just a few months of training, AlphaGo Zero, the latest version of AlphaGo, is far more capable than human experts. Beating an expert in chess (a complex game, but much less computative than Go) is a matter of hours.
Microsoft – the late horse
While Nvidia, Google, and Intel are all focused to some extent on AI, their chips provide service processing that takes place on the device, not in the cloud. Microsoft claims that the performance of its FPGAs cloud-based AI service is equal to or better than that of Nvidia, Google, and Intel. Microsoft believes its FPGA-based cloud solution, code-named Project Brainwave, will outperform cpus, Gpus and TPus in terms of scalability and flexibility.
In general, processor-based solutions are somewhat limited by design to accomplish specific tasks. However, the flexibility and reprogrammability of fpGas make upgrading easier and processor performance higher. On an Intel Stratix 10 FPGA, Microsoft’s Project Brainwave achieved 39.5 teraflops and a latency of less than a millisecond, according to Microsoft.
Whether FPGas can provide the best solution for ARTIFICIAL intelligence is as open to debate as any other issue. Microsoft argues that the production costs of creating AI-specific asics are too high, while others argue that FPgas will never fully achieve the performance of chips designed specifically for AI.
In a paper presented at the International Symposium on Field Programmable Gate Arrays (ISFPGA) in March, researchers at Intel’s Accelerator Architecture Lab Says it compared two generations of Intel FPGas (Arria10 and Stratix 10) and Nvidia Titan X Pascal (Titan V processor) that handle deep neural network algorithms. According to Intel researchers: “The results showed that Stratix 10 FPgas performed 10%, 50%, and 5.4 times better (TOP/second) than Titan X Pascal graphics processors for pruned, Int6, and binary DNNs matrix multiplication, respectively. In ternary-Resnet, Stratix 10 FPGas perform 60% better than Titan X Pascal Gpus, with a performance power ratio 2.3 times higher. This suggests that FPGas could be the platform of choice for accelerating the next generation of DNN. “
In China, although not as fierce as foreign countries, the power of start-ups can not be underestimated.
Cambrian released three smart chips this year: The Cambrian 1H8 for low-power scene vision applications, the Cambrian 1H16 with wider versatility and higher performance, and the Cambrian 1M for intelligent driving. In addition, there are more shenjian Technology, Kneron, Kunyun technology and other start-ups are also focusing on the research of smart chips, Horizon’s current situation can not be said to be insecure, besieged on all sides, but it is by no means occupy very favorable conditions.
Ambition on the horizon
Horizon’s ambitions may be better reflected in three sets of intelligent solutions beyond the chip.
Whether it is Nvidia’s GTC China 2017 conference this year or Baidu World Conference later, numerous enterprises have mentioned that they will start the layout of smart city. Today, Horizon has joined the team.
We mentioned at the beginning of the article that Horizon had already provided solutions in the field of intelligent driving as early as 2016. Dr. Kai Yu also said in his speech that the high requirements of accuracy, power consumption, real-time performance and reliability of automatic driving processors represent the Everest of ARTIFICIAL intelligence processors. With the release of its own smart chip, Horizon has a more sufficient foundation for the development of intelligent driving, smart city and other fields.
This new product release, let horizon return to people’s vision, and the newly released intelligent chip can get the recognition of the industry is still in the unknown state, is a stone to stir up thousands of waves, or as sand is generally swallowed by the sea? AI front will continue to pay attention to the latest developments of Horizon, whether Horizon has the absolute strength to occupy a position in the domestic chip market, we are very looking forward to the moment when the light shines from Horizon.
For a detailed analysis of the 2018 chip wars, please check out our third article “2018: AI Chip Wars will start”, and stay tuned for AI Front’s year-end smart chip feature!
Follow our wechat account “AI Front “and reply to “AI” in the background to obtain the SERIES of “AI Front “PDF e-books