Since there are already many self-developed AI chips, why should Baidu cooperate with Intel to develop the Nervana Neural Network Training Processor (NNP-T)?
Self-developed AI chips do not represent complete "self-sufficiency"
Baidu has released the AI chip for cloud and terminal. The latest release is Hongyuan's far-field voice interaction chip. The core capability is real-time processing of far-field array signals, high-precision ultra-low false alarm voice wake-up, and offline voice recognition. According to Baidu, this chip is built according to the car standard, which will bring great convenience for car voice interaction and smart furniture. Different from Hongjun's terminal, Kunlun is a cloud AI chip, including training chip Kunlun 818-300, and reasoning chip Kunlun 818-100, which will face smart cars, smart devices, voice images and other scenes.
It can be clarified that Baidu considers the smart car field when designing the cloud or the terminal AI chip. This combines the advantages of Baidu in the AI field and will also help Baidu's advantage in the field of automatic driving.
However, autonomous driving is only an important scene of AI landing, and AI will become ubiquitous in the future. Baidu naturally does not regard AI as a workload, limited to a few scenes such as autonomous vehicles or voice recognition.
Therefore, in the face of the different needs of different scenarios, different AI chips are needed, which are satisfied by various selection schemes. It can be seen that whether it is Baidu, Ali, Google or Amazon, many technology giants have entered the AI chip market in this round of AI boom. Technology giants hope to use their own research and development of chips to better explore the value of the vast amounts of data they already have, while reducing costs and maximizing their lead in some areas.
At the same time, technology giants still maintain demand for products such as CPUs, GPUs, and FPGAs that already exist and are mature in both hardware and software. Therefore, the independent development of AI chips by technology giants does not mean that traditional chip giants will be replaced. The technology giants are self-developing AI chips in the hope of playing a better role in specific areas.
How does Intel view this phenomenon as an important AI chip player in the world?Naveen Rao, Intel's vice president and general manager of the Artificial Intelligence Products Division, said in an interview with the media: "Many companies have entered the AI chip market, and in fact they have also reflected that demand for AI computing is very strong and high. We welcome and enjoy it. See the market with a variety of different solutions, especially in the next one to two years, this is the inevitable result of innovation. However, AI can not be provided by a single chip, but through A variety of options are available, each tailored to the specific application. The future of AI is a heterogeneous world that relies on deep collaboration between partners to generate better solutions and drive The full realization of AI value."
Why co-develop the Nervana neural network training processor?
At the 2019 Baidu AI Developers Conference, Naveen Rao announced that Intel is working with Baidu to develop the Intel Nervana Neural Network Training Processor (NNP-T). This collaboration includes a new custom accelerator to achieve the goal of speed training deep learning models. It is reported that NNP-T is a new type of highly developed deep learning system hardware that can accelerate large-scale distributed training.
Nervana is a deep learning startup that Intel acquired three years ago. CES 2019 Intel released the first Nervana series neural network processor NNP-I, which is suitable for the acceleration of enterprise-level high-load reasoning tasks. Intel plans to put into production in 2019. Intel also revealed that Facebook is a development partner for its NNP-I chip.
The new NNP-T, Intel also chose Baidu, an important partner for many years. Why is that?We look for answers from the collaborative development of NNP-T and the work of both parties.Naveen Rao said: "When working with Baidu to develop NNP-T, Intel's main contribution is to provide a new architecture for neural network processing. And, we have made some achievements in software, so that the entire ecosystem related partners can On top of our structure, they do their work. In cooperation, Baidu mainly provides feedback on actual use and training, including feedback they receive when serving customers."
Naveen Rao further said: "Usually people think that we are just doing a neural network calculation, but in fact our cooperation content is not limited to neural network computing. In addition to the calculation itself, we must also consider that we provide The serviceability and effectiveness of the solution, including macro standards, enable subsequent solutions to be expanded horizontally and vertically. In addition, we will closely monitor the specific problems, such as whether there is a problem with the connection, which A chip will not go wrong, etc."
It can be seen that the Nervana series is a dedicated neural network processor that achieves the best value for the best energy efficiency ratio for specific algorithms and domains. This will also explain why each of the Nervana series processors has a corresponding cooperative development partner.
Of course, this kind of cooperation can achieve a win-win situation.Intel has a deep accumulation in the chip field. Working with technology giants to develop dedicated AI chips will not only enhance their competitiveness in the AI era, but also better understand customer needs. For the technology giants who are cooperating to develop dedicated AI chips, the cooperation with Intel can reduce the difficulty of developing AI chips. With the cooperation, the customized chips can meet the AI needs of the technology giants more quickly and better. .
How is the AI chip competing in the future?
The demand for AI chips is diverse, and the relationship between traditional chip giants and technology giants will also become cooperative competition.However, the importance of both hardware and software in the AI era is becoming more and more important, and this will be one of the keys to future competition.Naveen Rao revealed that at present, the software-related part of his work will account for 60% to 70%, and there may be more in the future.
"For Intel, the reason why it will gradually increase its investment in software is also to notice the new trend of heterogeneous computing. The previous x86 architecture,MicrosoftWill develop an operating system running on the x86 architecture. But after the emergence of heterogeneous computing, the situation has changed. We often need to build that layer ourselves, so in the end there will be a question: Who is the management software? This thing is hard to define. Now Intel itself is very active in the work of some developer communities, providing developers with the ability to engage in development work, of course, development is based on our hardware. Nveen Rao said.
Therefore, the competition of AI era chips will require more diversified innovations, including heterogeneous computing, as well as soft and hard coordination. Quantum computing, neural mimetic calculations, and silicon light calculations are all directions for future calculations.In the frontier calculations, chip giants have an advantage. As early as 2017, Intel introduced the self-learning neural chip Loihi. In June last year, Intel said researchers were testing a tiny new "spin qubit" chip, which is smaller than the pencil eraser and is Intel's smallest quantum computing chip.
It is also worth mentioning that software and hardware are concentrated on an AI platform that can be further enhanced with more technology blessings, such as the higher memory performance provided by Intel AoTeng data center-level persistent memory, through which Baidu can The Feed Stream service provides personalized mobile content to millions of users and a more efficient customer experience through the Baidu AI recommendation engine.
In addition, given the importance of data security to users, Intel is working with Baidu to create MesaTEE, the Memory Security Feature as a Service (FaaS) computing framework based on Intel Software Protection Extensions (SGX) technology.
The demand for AI has attracted technology giants to compete in the AI chip market, but their entry does not mean that traditional chip giants will be replaced because AI's needs are diverse, requiring different AI chips and targeted solutions to meet demand. . Moreover, the entry of technology giants into this market does not mean ultimate success. In addition to the long-term, high-input, and large-scale professional talents of the chip industry, many small factors such as software and ecology determine its success or failure. Therefore, although there is more competition in specific areas, the future technology giants and traditional chip companies will still maintain cooperation and win-win.
Visit the purchase page: