[Intel Reveals First Commercial Neural Network Processor at First AI Developer Conference, VPU Performance Soars 3-4 Times]
This is the second time I have seen NaveenRao.
Like AI, he talked about AI and his words were opened up. There have been an endless stream of ideas and theories.
Intel Global Vice President and General Manager of Artificial Intelligence Products (AIPG) NaveenRao
It is his passion for extreme sports that forms an anti-eruption with his warm professor-like temperament. The 40-year-old AI expert was also a sportsman, so much so that he injured all fingers in skiing, skateboarding, cycling, driving racing, wrestling and playing basketball. Perhaps such an adventurer is more suitable for artificial intelligence. After the founder of his deep learning startup Nervana was acquired by Intel, Nervana was quickly integrated into the Intel AI core strategy. Now Rao has become Intel's artificial intelligence business group ( At the helm of AIPG).
Rao said that coming to Intel, "here is an open culture." He likes teamwork, but calling resources is not an easy task, but Intel has extensive experience in the marketization of products, and strong centripetal forces are The various departments of the company are twisted into one team and work together towards one goal.
At Intel, working hard is always more important than empty talk. At Intel's first AI developer conference, led by Rao led the team, Intel's artificial intelligence business related departments' "defective role" concentrated appearances, this should be a historical precedent for Intel AI. You know, in addition to internal Intel conferences, it is possible to publicly see a group of “big cows†that are so cleverly focused and almost zero-probability events.
But Intel did not disappoint.
This time it has come up with a super high-profile Intel "AI Galaxy" (let's call it), as shown below, from left to right:
Jennifer Huffstetler, Vice President and General Manager, Data Center Product and Storage Marketing, Intel Data Center Division
ReynetteAu, Vice President, Intel Programmable Solutions Group
JackWeast, Senior Principal Engineer and Chief Architect, Intel Driver Solutions
Gayle Sheppard, Vice President of Intel's New Technology Division and General Manager of Saffron's Artificial Intelligence Division
Remi El-Ouazzane, Vice President of Intel's New Technology Division and General Manager of Moviduis
Jonathon Ballon, Vice President, Intel Internet of Things Division
NaveenRao, Vice President of Intel Corporation and General Manager of Artificial Intelligence Products Division
Although this lineup is comparable to Marvel's "Avengers' Alliance", there is still a "Great God" outside the picture frame.
Carey Kloss, Vice President of Intel's Artificial Intelligence Division and Core Member of the Nervana Team
Carey Kloss is vice president of the Intel Artificial Intelligence Business Group and a core member of the Nervana team. Although he did not appear in the above figure, he expressed love for the team to the tech traveler. "Intel has the best post-silicon I've seen so far. Post-silicon-bring-up and architecture analysis.†Because of this, the Nervana Neural Network Processor (NNP) has been greatly improved.
In fact, NNP is Intel's brewing long-term "killer." At this AI Developer Conference, Rao quickly revealed Intel's next-generation AI core, the Intel Nervana NNP-L1000, code-named "SpringCrest," a dedicated artificial intelligence chip, and this chip is about to become Intel's first commercial nerve. Network processor products are planned for release in 2019.
Although Rao did not disclose the details of the new generation of AI chips, Carey Kloss, who is also a founding member of Nervana, holds secrets - we certainly won't let him go. During the AI ​​developer conference, the tech traveler conducted a "grounding gas" dialogue with him, and Intel, who was originally known as "wishful thinking", could play like this.
NervanaNNP: New AI core performance soars 3-4 times, but the power has not yet fully released
In Rao's one-hour keynote speech, the most important release of non-Intel Nervana neural network processor is none Other than its significance to Intel.
If you use the "LakeCrest" (NervanaNNP series initial chip code) announced for the first time in October last year, you can say that "LakeCrest" is like a "timely rain" and successfully helped Intel stand in the ASIC chip competition. However, Intel has proposed a bigger goal, that by 2020 to enhance the performance of deep learning training 100 times. The Crest family is likely to become the fastest way to achieve Intel's goals.
To know that the creation of a chip is not easy, if there is not a crazy, dedicated chip development team behind it, it will also be an unsatisfactory chip. Therefore, the problem that the insider who understands the doorway is more focused is: How does the Intel IC design team behind the Nervana neural network processor series chip create the "SpringCrest" that can soar to 3-4 times the existing performance?
Although Carey Kloss has a tight-lipped tone, but with regard to the Nervana neural network processor, the tech traveler is still in the chat with him, get the following sharp information:
1. The main differences between LakeCrest and SpringCrest
LakeCrest, as the first generation processor, achieves very good computational utilization on GEMM (matrix operations) and convolutional nerves. This is not just the utilization of 96% throughput, but in the absence of full customization, Nervana also achieves a computational utilization rate of more than 80% for GEMM in most cases. When developing next-generation chips, if they can maintain high computational utilisation, the new products will have three to four times the performance improvement.
2. The utilization rate of LakeCrest has reached 96%. Why did SpringCrest not rise instead of rising?
This is a market strategy that will reduce the utilization rate appropriately. Some cases can indeed achieve 98%, and in the absence of resource conflicts, each silicon chip can operate at 99% or even 100% computational utilization. But what Intel wants to show is the utilization rate that can be achieved in most cases, so it has been properly adjusted.
3. Why did the Nervana chip's release rhythm continue to be delayed?
In two phases, Nervana began researching and developing LakeCrest at the beginning of 2014. At that time, the entire team was about 45 people and was building the largest Die (silicon chip). We developed Neon (deep learning software) and also built Cloud stacks, these are all done by small teams. But this is also a challenge. The growth of small teams will be painful. It took Nervana a long time to bring out the first batch of products. Until last year, the chip was actually introduced. About why SpringCrest chose to launch at the end of 2019, because it needs to integrate more Die (silicon chips) to get faster processing speed, but it takes a certain amount of time to manufacture silicon wafers and it also needs to become a new neural network processor. This is the reason for the delay. Currently, SpringCrest is at a reasonable pace and has all the ingredients for success next year.
4. What are the adverse effects of delay on Intel?
CareyKloss does not believe that Intel will be at a disadvantage in the neural network processor, because Intel's response speed is relatively fast, for example, gradually turning to bfloat is an important factor, it is a widely used numerical network data format for the neural network, It is very popular with the market. In the future, Intel will expand support for bfloat16 on artificial intelligence products, including Xeon processors and FPGAs.
5, compared nGraph with CUDA: not afraid
Regardless of the hardware level, Intel has also increased its software deployment. Currently, the Intel AIPG Division is developing software called nGraph, a framework-neutral deep neural network (DNN) model compiler. Intel is integrating deep learning frameworks such as TensorFlow, MXNet, PaddlePaddle, CNTK, and ONNX on top of nGraph.
It is also a platform concept. Many people like to use the GPU to represent the company Nvidia and Intel. In fact, Carey Kloss bluntly stated the difference between nGraph and the competitor CUDA platform.
“nGraph is not the same as CUDA. CUDA can be understood as the bottom surface of nGraph. We call it a transformer. nGraph receives input from TensorFlow, Caffe or MXNet via a fixed API, and then optimizes the performance through the graphics compiler. Something that is not needed is then sent to the CPU's MKL-DNN, so the CPU still uses MKL-DNN, even in nGraph. "It's easy to see that Intel is also interested in putting chip development on a unified platform, nGraph builds on the interface to develop AI applications based on all Intel chips.
Compared to the new generation of Nervana NNP-L1000 is still in the R & D stage, Intel's other chip VPU focused on computer vision is actually commercially available. Regarding this chip, Intel also pinned what kind of market expectation, and came to see the answer of another Great God who was also outside the picture frame.
Movidius VPU: May Be a Killer Application in Computer Vision, Including Windows 10
Gary Brown is Marketing Director of Intel Movidius. His main job is to buy products developed by Movidius to a good market and make the sale bigger. He told technologist, "anything related to computer vision and camera can use Movidius".
Intel's Movidius Marketing Director Gary Brown
The chip developed by Movidius is called visual processing unit VPU. It is a chip that combines both computer vision and smart camera processor. The processing is divided into three categories: The first type is ISP (image signal) processing, the second is Classes are based on camera capture technology and the third category is computer vision and deep learning. According to Gary Brown, VPU is currently "working" in VR products, robotics, smart homes, industrial cameras and AI cameras, surveillance and security.
The popularity is undoubtedly the two advantages VPU has: First, it can run neural network directly on the local camera, and it has lower delay and less electricity consumption than sending data to the cloud and sending it back to the local. The shorter time also means lower bandwidth and cost; the other is the energy-saving technology. With the front-end algorithm to reduce power consumption, most chips can be turned off, and only a small part of the optimized face detection function is operated. When the face appeared, other chips were activated, so that the face monitoring system could be kept on. There was no problem for the home camera for 6 months. Movidius is also adding a neural computing engine to the latest product MyriadX's VPU, which can increase the chip's performance in deep neural networks by a factor of 10.
"Snapshot artifacts" Google Clips camera is another typical application of VPU. This camera, which is small and even has no screen, can record the picture “automaticallyâ€. In reality, the VPU is playing a role. Driving the camera with AI sounds cool, but it's just the tip of the iceberg for the VPU application. Because Intel also hopes for a larger "business" for VPU, far beyond the scope of the hardware, attaches great importance to cooperation with the software giant Microsoft Windows10.
"Windows10 may soon become one of the huge markets for Movidius VPU." Gary Brown became excited. "Microsoft recently developed a new API for Windows 10, called WindowsML, which stands for machine learning. Everyone can write applications for Windows 10 and transfer machine learning to Movidius VPU." Developers can use WindowsML for application development, such as visual applications, video conferencing applications, smart assistants for image search, and search for interesting things through image recognition.†That is, if Windows ML is used in Windows 10, it means There is no need to run machine learning on the CPU.
Gary Brown also said that there are currently PC manufacturers and he negotiate to put the VPU directly into the computer's new module, but the name is not convenient to disclose.
In addition to the chip's reputation in the market, it was surprising that Movidius's other strength turned out to be software. Because "many members of the Movidius team belong to the development group, hardware is only part of our product, and the software developer kit contains libraries, drivers, open source and corresponding firmware, and is also one of the Movidius product lines." Gary Brown also stated that the new version of Intel's software The developer tool is OpenVINO. The toolkit helps developers create and train AI models in the cloud (such as TensorFlow, MXNet and Caffe, and other popular frameworks) and deploy them to various products like Movidius and Hikvision. Cooperation is to adopt this model.
At present, 75% of Movidius' customers are concentrated in the Internet of Things. This is not unusual. If you know a little about Movidius, you know that this computer vision startup was first acquired by the Intel IoT department a year and a half ago. In order to export more comprehensive AI capabilities, Intel’s internal departments are now intertwined with webs. The relationship, including the penetration of AI in the Internet of Things, is one of Intel’s most important businesses. How can this “combination card†be used? Another big cow debuted.
Intel AIxIoT: Focus on "smart" objects, not just computing power
Jonathon Ballon, vice president of Intel's IoT business unit, who specializes in induction, threw out three general conclusions about the content of the Internet of Things at the opening:
There is no universal framework for the Internet of Things. There are many kinds of architectures according to different scenarios.
Not all artificial intelligence occurs in the data center or the cloud. Artificial intelligence runs in a distributed computing architecture: from cloud, network, and edge devices.
Intel has made great investments in software tools. Intel believes that software is an important factor in hardware platform differences. The complementarity of the two tools, nGraph and OpenVINO, enables heterogeneous architectures to perform optimally.
Jonathon Ballon, Vice President, Intel IoT Business Unit
Concise and concise, as Jonathon Ballon said, the Internet of Things does not have a panacea, but how do you hit it? Intel seems to have unique thinking and roads.
"We focus on smart objects, not just computing power." Jonathon Ballon further explained that "a device with a chip, with computing power, this is called a computing device, but this does not mean that it is smart. When put The computing device connects to the network and separates the data. At this time, the device is called an IoT device, but it is not necessarily a smart device. The difference between the IoT device and the smart device is that the latter has the learning ability. Where artificial intelligence can play a big role."
If this thinking is mapped to the field of medical imaging, Jonathon Ballon also talked about the trajectory of “smart†evolution: “Perhaps we have focused on rapid image acquisition over the past decade, but now that AI development has crossed this stage, we are thinking How to analyze impacts more quickly and accurately than doctors, allow the AI ​​system to process 10,000 medical images in a matter of minutes, and indicate to radiologists which images should be of special interest. The next step is to associate them with related cases. Diagnosis and treatment plan."
The same disruption will also occur in the retail market: D-MART “Don’t Shop†built by Intel and JD has been deployed in multiple smart stores and smart vending machines. The machine learning algorithms used by unattended stores are mainly concentrated in the three directions of knowledge, knowledge, and knowledge. As the data involved in online and offline communication is turned on, unstructured data such as video is converted into structural data, etc., which needs to be used in the field of machine vision. CNN (Convolutional Neural Network) algorithm, traditional machine learning algorithms used in the intelligent supply chain, such as SVM, linear regression of statistics, and logistic regression. Considered comprehensively, Jingdong chose Intel's edge server to do hardware layer support.
From medical to retail, we only see the microcosm of IoT practice. In fact, Intel is trying a general logic or methodology to make the deployment of the Internet of Things open in all industries. According to Jonathon Ballon, the Internet of Things also has its own “Moore's Lawâ€. IOTs in different industries will experience the same three phases: connectivity, intelligence, and autonomy. The autonomy phase is the ultimate trend in the future development of the Internet of Things. The "no shop" built by Intel, JD.com, and Amazon is an example of completely independent operation.
When it comes to competition, Jonathon Ballon said he was not worried. "Intel's strength is to make good use of distributed computing architectures and focus on building end-to-end solutions that include devices, gateways, networks, clouds, data centers, etc." As you can see, not only AI's product portfolio, but also the IoT, Intel has also made a good "combination card". From the equipment to the cloud, it has prepared a complete set of product portfolio solutions.
So, what are the cattle X cards in the hands of Intel?
Intel never directly tells others how strong its AI capabilities are, but the AI ​​around you may have "Intelinside" in the future.
Just like you haven't heard of Movidius VPU, but you may know that DJI launched a mini drone with gesture recognition remote control this year; not familiar with the name Moblieye, but you have probably heard of the autopilot feature of Tesla Autopilot; I haven't studied Intel AI platform, but you may be surprised by the live broadcast of the 2020 Tokyo Olympic Games; even an Intel brand is added to the chips used by most of today's artificial intelligence hardware companies.
In fact, by virtue of its size, Intel has completed a far-reaching layout of AI hardware. From training to reasoning, there are intelinside figures from the server to the terminal AI's entire industry chain. If data flooding presents great opportunities and challenges, a wide variety of application needs require different solutions and technologies to meet, and the same applies to artificial intelligence. Complex workloads also require different types and characteristics of artificial intelligence products to support, which requires a more comprehensive enterprise-level solution.
In terms of artificial intelligence strategy, Intel actually has always emphasized "breadth", that is, for each architecture style, Intel has one or more product portfolios, so that organizations of all sizes can open their own labors through Intel. Smart R&D. For example, Intel is working with Novartis to use deep neural networks to accelerate high content screening - a critical element in early drug development. The cooperation between the two sides shortened the time for training image analysis models from 11 hours to 31 minutes.
Of course, to release AI potential, only "breadth" is not enough, and there should be more comprehensive considerations. After more than a year of combing and consolidation, Intel finally came up with a relatively complete product portfolio - Intel's full-stack solution for artificial intelligence, including Xeon Extensible Processors, Intel Nervana Neural Network Processors and FPGAs. , network and storage technologies; Intel architecture-based math library (IntelMKL) and data analysis acceleration library (IntelDAAL) optimized for deep learning/machine learning; support and optimization of open source deep learning frameworks such as Spark, Caffe, Theano, and Neon et al.; built a platform represented by Intel Movidius and Saffron to promote front-end and back-end collaborative artificial intelligence.
Is this enough? of course not.
Almost missed "Loihi". It is a neuroscience computational chip that Intel is developing. It can transmit information through pulses or spikes like the brain and calculate through “asynchronous activation†method, which makes machine learning more efficient and at the same time requires less computational power. However, Rao also pointed out that Loihi is currently only a research project and is an important research direction of Intel but not the only direction.
"If we can improve, this technology will become a potential stock."
"At the same time, there is quantum computing, which is a way to create more computing power."
Ecome mobile radio,Portable Two Way Radio,Motorola Vhf Handheld Radio,Multi Band Portable Radio
Guangzhou Etmy Technology Co., Ltd. , https://www.digitaltalkie.com