AI Frontline Reading: “In 2017, the AI ​​chip is the highlight of the semiconductor industry, and it has attracted far more attention than the semiconductor industry. This year, from the tech giant to the start-up company, the new and old roles took turns and we staged for us. It's a wonderful show. After a few years, when we look back, we can definitely use 2017 as the first year of the AI ​​chip."
The "Dislocation War" between GoolevsNvidia Giants
In early April, Google announced a paper to be published at ISCA2017: "In-DatacenterPerformanceAnalysisofaTensorProcessingUnit". It can be said that it is this "little thing" that unveiled the curtain of an annual drama, and the far-reaching effects it may have may even last many years later. In fact, in June 2016, Google disclosed that it had developed a dedicated AI chip for use in the cloud, the TPU (Tensor Processing Unit). Google's AI chip is certainly eye-catching news, but it suffers from not publishing details, and everyone can only guess and wait. Therefore, this common academic paper received great media attention. I also wrote a review article for the first time: "GoogleTPU reveals secrets," and it is one of my articles with the most public number reading. Of particular concern to TPUs is not only our melons, but also Nvidia, the absolute ruler in the AI ​​chip field. Later, a verbal battle between the Yellow Master and Google regarding the TPU Benchmark results was justified. As early as 2016, when Google revealed TPU, Nvidia repeatedly stated that it had no threat to the GPU's dominance in AI computing.
On May 11th, at the Nvidia GTC 2017 Conference, the Yellow Master throws out the latest GPU Volta (GV100) on Keynote. Nvidia stocks rose sharply, and the media also reported it. The focus of the AI ​​chip seems to have returned to the Nvidia side.
In addition to the announcement of the heavyweight Volta, there is a "little incident" on the GTC. Nvidia announced that it will open its DeepLearning Accelerator (DLA), which will be officially released in September. This announcement was brought in a word in the Yellow Key's Keynote, but the shock caused by the industry is not small at all. "Why do Nvidia engage in open source? What will be open source? Will this open source affect the prospects of many startups?" Discussions on these issues have continued until NVDLA is truly open source.
Before long, on May 17th, at the Google I/O Conference, Google announced the second-generation TPU, using the words “...stoleNvidia'srecentVoltaGPUthunder...†in the media. Although the details of TPU2 have not been announced in detail, the indicators do indeed look good and have very good scalability. The only regret is that it does not sell externally. It can only be used by everyone in the way of TPUCloud.
In late September, Jeff Dean, Google's great software god, participated in the HotChip chip conference. He also personally introduced the situation of TPU and TPU2 in Keynote "Recent Advances ArtificialIntelligencevia MachineLearningandtheImplicationsforComputerSystemDesign" as an important part of the new computing ecosystem.
At the end of September, NVDLA open-sourced some of NVDLA's hardware code before the promised deadline. At the same time, it announced a road map for more open source hardware and software in the future. After that, everyone also made various analyses and discussions on NVDLA and tried to play it. From the current point of view, the open source of NVDLA does not seem to have affected the financing of many start-up companies. We will talk about this topic later. As for the reasons for Nvidia's open source DLA, the official statement is to make it easier for more people to implement Inference and promote the promotion of AI, especially on a large number of embedded devices. However, from the perspective of the entire open source process, this open source decision seems to have been rushed. DLA comes from a module in the Nvidia Automated Driving SoC and was not originally designed for open source IP. Moreover, the open source in September only disclosed a part of the hardware code and the corresponding verification environment. There is still a big gap from the real use. We are not sure whether this open source decision is related to the strong debut of Google TPU (which has a large advantage in Inference). But the basic assumption is that the core benefit of Nvidia in DeepLearning should lie in Training (the current GPU is still the best platform for training). It is in its best interest to let Inference be lower and penetrate into more applications, especially Edge, to further promote Training's needs. And NVDLA's software environment is still using Nvidia's CUDA/TensorRT, still controlled by Nvidia.
This began with a dissertation that spanned almost all of Google’s and Nvidia’s tussles in 2017. The impact on the industry may be far greater than the two companies themselves. I call it "misaligned" war because it happened between traditional software giants like Google and chip giants like Nvidia. If you switch to IntelvsNvidia, it seems to be normal. Google's participation in the war may be opening a new era. We can see that not only TPU, Google also announced in October their custom SoCIPU (ImageProcessingUnit) used in the "GooglePixel2" mobile phone. Like Apple's more and more custom chips, technology giants such as Google have applications (knowing exactly what they want), technology (the accumulation of related technologies over the years), resources (no shortage of money, no shortage of people). Advantages, customize your hardware, and even the chip will become normal. At the same time, we also saw that the demonstration effect of GoogleTPU has already appeared, and more tech giants have joined AI to accelerate hardware competition. Tesla announced its own customized autopilot chip; Amazon, Microsoft, and domestic BAT, Huawei provide dedicated FPGA acceleration support in the Cloud; BigFive is also said to have its own development of the chip; BAT is also building a chip design Team, etc. Although everyone's specific architecture and implementation are different, they all reflect great interest in AI-specific hardware. I believe this trend will become more apparent in the future.
At the same time, the traditional chip giant will certainly not sit in the grip of Nvidia or be divided by Google. Intel has acquired Nervana (cloud), Movidius (end-to-end), Mobileye (autopilot), Altera (FPGA) and AMD RajaKudori (GPU), and even Loihi (nueromorphic). A good hand; Although the action is not as fast as everyone thinks, but behind the force is still worth the wait. AMD is also trying to catch up, after all, their CPU + GPU has its own unique skills, and the entire company has gradually come out of the trough. Moreover, regardless of whether Tesla and AMD are cooperating with the autopilot chip in the end, the chip company's output chip design capability model is also a good (or frustrating) choice.
“The new computing model represented by DeepLearning will lead the development direction of the future chips,†which is basically a consensus for everyone. More and more players will focus on chips that can support new types of computing, many of which may not have been in the semiconductor industry before and do not understand what the chip is all about. In 2017, we can see some popular science articles comparing CPU, GPU, FPGA and ASIC architecture from time to time. Even with 10W+ readings, it is easy to see everyone's enthusiasm.
A long list of startups
In the 2017 AI chip drama, the protagonist is not only a giant, but also the start-up companies have made their appearances. More importantly, in the "performances" of start-up companies, Chinese companies are not only unrestrained but also very colorful. I started to maintain a list of AI chips on github from August, including the products of big companies and startups. By December, more and more information was on the list, and there were more than 30 start-up companies worldwide. And this list contains only public information, there are many companies in the stealth state and not included. I also heard a claim that there may be more than 100 start-up companies in the AI ​​chip field, and 30 in the TSMC queue.
No matter what the field, start-ups will face many risks and uncertainties, and they may also be constantly adjusted and changed during their growth. AI chips are certainly no exception. We have seen that in the course of this year, many companies have continued to grow, gradually clarifying their own direction and positioning, and they are getting more and more solid. On the other hand, from the perspective of this year's start-up company financing, there are obvious bubbles in this area (including the broader concept of AI). Some companies can achieve "PPT financing" or "Paper financing" without any real thing. Some companies focus on the PR. Kung Fu is done for VCs. It is called “2VC†company. In the face of AI's trending opportunities, bubbles are certainly normal, but it is hoped that these bubbles will not hurt the development of the entire market.
Aside from all kinds of smoke and bubbles, we gradually see some "leading companies" in this area. For example, the domestic Cambrian, Horizon, Shen Jian Technology, and Bit Continent all released their products in 2017; the US’s Cerebras, WaveComputing, Graphcore, and Groq (formerly the former Google TPU designer) have strong capabilities. Or have their own unique technologies and relatively clear products. In 2017, there are also some AI start-up companies that rely on application development chips. Most of these companies take the lead in research and development of chips. I also expect to see more of this in 2018. Of course, many start-up companies do not disclose their own information, and they do not rule out the possibility of a big move.
Anyone familiar with the semiconductor industry may be more aware that it is very difficult for VC startups in the semiconductor industry to invest in VCs. The main reason is that this industry is high risk, high threshold and long cycle. However, in 2017, startups of AI chips were sought after by funds. We can look at some of the open financing data for this year. Cambrian: US$100 million (approximate valuation of US$1 billion); Shen Jian Technology: US$40 million; Horizon: nearly US$100 million;; Cerabras: US$60 million (estimated US$860 million); Graphcore: US$50 million. I also mentioned earlier that when Nvidia announced that it wanted to open source DLA, everyone felt that it would have an impact on the financing and valuation of startups. However, from the results, this situation did not appear. After September, we saw many successful start-up companies financing. The enthusiasm of investors seems to have not weakened at all. As long as there is a new company, many investment institutions will immediately rush in.
Why do investors who have traditionally been reluctant to touch the semiconductor industry now rush toward AI chips? This is an interesting question. The specific reasons may have many aspects. The investment boom in the entire AI field should be one of the main reasons. If we look at the capital behind these investments, we can see that many investors themselves are active investors in the AI ​​field, and even themselves are tech giants that focus on AI as the future, such as BAT. The traditional investment in semiconductors is rather more cautious. From this perspective, these capitals, which do not have much semiconductor background, will enter the chip field in large numbers, which will bring new opportunities and horizons to everyone, or bring risks and uncertainties. It remains to be seen. In addition, the AI ​​chip now referred to generally refers to the DeepLearning acceleration chip. Relatively speaking, the key algorithm is simple and clear, and the optimization goal is very clear. Many technologies (such as hardware acceleration of matrix operations) have been researched for many years. It is relatively easy to verify, test and debug this kind of hardware accelerator. Without careful optimization, the hardware part can be completed by a smaller team in a shorter time. These technical features are more suitable for quick start-ups of startups. Of course, making hardware for an accelerated chip (or IP) is only the first step. To really make a product that can be accepted by the market, you need a lot of solid work, product definition, hardware performance, software tools, system testing, on-site support, etc. A short board can't be. Although everyone is very concerned about the time spent on filming, after the samples have been released, there are still more dirty jobs.
2018 What to watch
For 2018, I still look forward to it. As an engineer who has been engaged in the design of chip architecture for many years, I first expect to see some technical innovations. In 2017, I wrote a lot of articles on the analysis of AI chip-related technologies. By the end of the year, I was almost aesthetically fatigued (I believe the same is true for readers). It seems that there are fewer and fewer new things. At the end of 2017, there was a start-up company called Vathys, which opened several brain holes at once, a fully customizable AsynchronousLogic, an equivalent clock up to 12GHz (28nm process), High-density SRAM (1T-SRAM), on-chip memory capacity Can reach 1.5GB (28nm); Wireless3DStacking, 10,000GBit/S@~8fJ/bit. These technologies are either still at the stage of academic research or have been short-lived. Is it possible for a startup company to make these big moves at once? So when Vathys's boss sent an e-mail saying that they should add their company to my list of AI chips, I started to decline. However, from another perspective, even if they are completely fooling, it can be regarded as a pain point hit DeepLearning processor. Moreover, these technologies are also currently being studied. Perhaps they can come out with certain enthusiasm and huge financial support from the AI. So, I still hope to see them or other teams able to make breakthroughs in these several technologies, let us really excited. When it comes to technological breakthroughs, we can expect to see breakthroughs in storage technology and architectural innovations driven by new storage technologies, including Neuromorphic, in the future (and perhaps even further than in 2018).
Next, of course, is the giants' next move. Will Google's TPU sell itself to users other than its own and compete directly with Nvidia? At present, the ONNX camp has formed a confrontation with Google. Google, as the most complete manufacturer of the ecosystem, is very meaningful to promote TPU to consolidate its leading position. Which of BigFive and BAT will learn Google’s example directly from the chip? Will Alibaba's chip research start with AI? Can Intel not fully explode as everyone expects? How will Nvidia deal with the challenges from all parties and will it make more specialized acceleration chips instead of just adding TensorCore to the GPU? When does Qualcomm add hardware accelerators to mobile phone chips? What will ARM do next and will it swamp the embedded end? ... Just think about it, there will be many things worth looking forward to. Recently we have also seen that AMD and Intel have rarely announced cooperation in order to fight Nvidia. IBM is deeply involved with Nvidia on Power9. In 2018, we may still see more vertical integration between the industry giants.
The fate of startups is also the biggest attraction in 2018. I said in an earlier article that "For the startup of AI chips, even if it is not a graduation exam in 2018, at least it has reached the end of the semester...". In 2018, most start-up companies will submit the results of the first test (chips) and will also begin small-scale trials. I believe that there will be fairly fair Benchmarking results, and the "theoretical" indicators will be replaced by the actual "run points." Although for start-ups, mistakes can be tolerated, the first-generation chips cannot fully represent the future of the company. However, chip-making requires constant support from huge resources, and leaving the team at this stage can be very dangerous. Of course, the first phase out is also the best opportunity for truly outstanding companies. I am very much looking forward to seeing my classmates who can stand out in the exams and move on to a new level (or directly graduate); or, there will be faces that we do not know and suddenly come out. In addition, in 2018, there will be more traditional chip manufacturers in the edge of Edge to join the competition, Samsung, Qualcomm, MTK, Spreadtrum and so on; and ARM with absolute advantages in the embedded IP should also have greater action, these All may have a major impact on the fate of startup companies.
In the end, it is possible to change. Overall speaking, how AI as a whole will develop in 2018 is a matter of great concern to everyone. Will it continue to grow at a high rate, or will it develop steadily, or will it encounter problems that will go high? In either case, the AI ​​chip is bound to be affected by the general trend. What is more special is that the chip development cycle is about 9 to 18 months, which is much longer than the development and update cycles of software applications. Coupled with some hysteresis, the development of the chip is difficult to synchronize with the development of algorithms and applications. One of the more frightening problems in chip development is uncertainty in the future. In contrast, a predictable and stable growth environment is most conducive to chip R&D, allowing chip designers to better plan products and coordinate resources.
Another variation is that there is a huge change at the algorithm level, which is technical uncertainty. The most successful AI algorithm in recent years is based on deep learning of neural networks. This is the basis for the current demand for AI chips. It also determines that most AI chips now aim at accelerating this type of algorithm. If the basic algorithm requirements change, it will have a great impact on the design of the AI ​​chip. For example, there is a certain application-based low-precision network, which is to use very low accuracy in inference, even directly using binary networks. If this Inference is widely used, the current chip architecture may need to be reconsidered. For another example, if the Hinton God's capsule network is useful, it may also need a new chip architecture to support it. After all, the AI ​​field is rapidly developing, so everyone must also keep a close eye on the latest advances in the application and algorithmic aspects. We also have to ask ourselves the following questions at any time (from Jeff Dean's speech at NIPS2017).
to sum up
2017 will soon be over. In the relatively "dull" semi-conductor field in recent years, the AI ​​chip has given us a little excitement. In fact, there are many things that can be talked about. The above text is basically where you think, and it is also a bit of an individual's sentiment. Accurate places, but also please correct me a lot.
I wish you all the best in 2018! I would like to wish all of you my colleagues in the first line of AI chips to succeed!
Pressure Thermometer,Capillary Thermometer Measured,Capillary Thermometer Measured In Boiler,Stainless Steel Pressure Gauge
ZHOUSHAN JIAERLING METER CO.,LTD , https://www.zsjrlmeter.com