Intel has announced a significant advance for its neuromorphic research processor, named Codehi. The company has now extended its implementation of Loihi to 64 processors, allowing it to create a system with over 8 million neurons (8.3 million). This new configuration (codenamed Pohoiki Beach) delivers 1000 times better performance than conventional processors in applications such as sparse coding, graph search, and constraint satisfaction problems. Intel says the new Pohoiki range offers 10,000 times more energy efficiency than conventional processor architectures in this type of test.
Neuromorphic Computing is a subset of the computer that attempts to imitate the architecture of the brain with the help of modern technological analogues. Instead of implementing a typical CPU clock, for example. Loihi is based on a peaked neural network architecture. The basic Loihi processor contains 128 neuromorphic cores, three Lakefield processor cores (Intel Quark) and an off-chip communications network. Theoretically, Loihi can handle up to 4,096 on-chip cores and 16,384 chips, although Intel has announced that it does not intend to market a design of this size.
"Thanks to the Loihi chip, we were able to demonstrate a power consumption 109 times lower than that of a real-time graphics processor, and a power consumption 5 times lower than that of an inference material. IoT specialized, "said Chris Eliasmith. CEO of Applied Brain Research and Professor at the University of Waterloo. "Even better, while the network is multiplied by 50, Loihi retains performance results in real time and uses 30% more power, while the IoT hardware consumes 500% more power and is no longer in time. real."
The implementation of Pohoiki Beach is not the largest deployment planned for the neuromorphic chip. Intel announces its intention to roll out an even larger concept, dubbed Pohoiki Springs, which will offer "unprecedented levels of performance and efficiency for enhanced neuromorphic workloads."
We covered the advances and research in neuromorphic computing for several years at ET. The work done on these processors is closely related to the work done in AI and the machine intelligence in general, but it's not just about how to perform AI / ML workloads. on existing chips. The ultimate goal is to build processors closer to the human brain.
One of the quirks of computing is that the analogies between the functioning of the human brain and the operation of computers are widespread. Human brains and conventional computers do not overlap very little on their functioning. Transistors are not equivalent to neurons and the pulsed neural network that Loihi uses to transmit information about his own processor cores is intended to be closer to biological processes that humans and other animals use than traditional silicon.
Projects like this one have several long-term research goals, but one of the most fundamental is to better understand how the brain works to replicate some of their energy efficiency. The human brain operates at around 20W. The Exascale supercomputer, considered the minimum for advanced neural simulation of anything more complex than an earthworm, should consume megawatts of power per supercomputer. The difference between these numbers explains why we are primarily interested in the long-term energy efficiency and computing potential of the brain. Architectures such as Loihi are not just an effort to write programs that mimic what humans can do. the goal is to copy aspects of our neurology as well. This makes their progress a little more interesting.
Background image: Tim Herman / Intel Corporation
At Semicon West 2019, a panel of industry experts debated whether Moore's law – the big prediction provided by Gordon Moore, said the number of components per integrated circuit would double over a period predictable time (originally 12 months, later extended to 24 months) – was still alive. Over the past decade, whether Moore's Law was viable in the long run or whether it was already dead and replaced by other methods of scaling up has become more commonplace.
Not surprisingly, there are many points of view on the subject. I have discussed in the past that if the text of Gordon Moore's prediction of 1965 did not change, how did people including it has a lot. Almost from the beginning, Moore's law – which initially provided for the density of transistors – was expanded to include performance predictions. The concept of scaling nodes has also changed over the years; It's been 20 years since node names refer to a specific entity size. When TSMC, Samsung, or Intel talk about a new node, they mean they have combined a combination of new technical approaches, smaller feature sizes (in some cases), significant changes, and manufacturing enhancements that Collectively justify have developed a new way of manufacturing transistors.
This kind of definition leaves room for a lot of play. According to Aart de Geus, co-CEO of Synopsys:
Moore's Law is based on the behavior of an exponential who has a technical back to the exponential that has revolutionized what humanity can do. The reason I say it 's totally alive, is that we are facing an extra one or two decades of incredible opportunities that will economically boost the thrust of technology without stopping. Maybe it is not exactly the same curve as that drawn by Moore, it does not matter.
Victor Peng, CEO of Xilinx, singled out the traditional scaling up of Moore's Law that offered advantages in terms of performance, power, and surface (that's Moore's Law + Dennard Scaling , and the fact that the CEO has handled the issue in this way testifies to the concepts have been mixed over the years.According to him, when it was possible to see benefits in these three areas over time, now companies had to focus on one or two of them and in that sense Moore's law was no longer working.
Lisa Su, CEO of AMD, who recently launched the most revolutionary processors his company has opened the door in 15 years, does not see that Moore's law is disappearing, but think that it is slowing down and that companies need to integrate more and more technologies outside the traditional field of the scaling the nodes to succeed. Other industry stakeholders shared their own points of view (Semi Engineering see you). In your opinion, the evolution of Moore's Law depends on the other technologies used by your industry to continue to advance silicon, be it NAND 3D stacking, advanced packaging techniques, or new hardware technologies. offering better performance.
Yet, despite the wide debate over whether Moore's law can be properly described as "dead," I think the very fact that we are discussing this point is itself instructive. If you have been a computer enthusiast for quite a long time, you remember a time when even asking the question "Is Moore's Law Dead?" You would have brought a funny look. Whether Moore's Law mentions transistor densities or that it also includes aspects of performance and power, these are factors that people have been debating and confusing for decades. The fact that it existed and worked was, however, inviolable. Today, we see people who are openly looking for other methods of definition that allow Moore's Law to Carry on to be true, mainly because of the classic definition no longer works in fact.
But once you have reached the point where the old definition actually works, you have arrived at a transition point. It is possible to retain the concept of Moore's Law to describe the equivalent improvement, but we have thus changed the fundamental nature of what Moore's Law was supposed to be. In the end, Gordon Moore's foundational article becomes a sort of ship of Theseus's problem – when is Moore's law itself no longer Moore's law? When did we dissociate the technical reality from what was originally described so perfectly from what the law supposedly refers to who has more to do with it? ;other?
It would be quite amusing if, in 500 years, AI-based hard light structures, which serve as frames for large conglomerates made up of human and business firms, are still talking gleefully about Moore's Law advances, describing improvements to the latest quantum peanut butter. Computer singularity – based on Gordon Moore technology that I have never heard of and circuit designs that I could not imagine.
Given how marketing works, I would not bet against that.
If you have not heard, the United States and China are engaged in a trade war. On the American side, one of the major problems is the problem of forced partnerships and intellectual property theft that US companies often report the price of doing business in China. China uses many sophisticated technologies, but according to the chip designers themselves, it will be impossible to meet Beijing's ambitious targets for local production if the United States refuses to cooperate.
There are alternatives in China, but the technological gap is too big, "said one of the leaders of one of China's leading artificial intelligence chip manufacturers, which relies on United States technology for chip design, said Nikkei "If we lose access to the US software or can no longer receive updates, our chip development will end up in a dead end."
Officials from NextVPU, an AI company created by former AMD employees, also echoed his remarks, including a representative of Semiconductor Manufacturing International Co, or SMIC, the most important foundry in mainland China. "We would use all the equipment and chip materials we have locally if their performance was sufficient," said a senior official at Semiconductor Manufacturing International Co., the largest chip manufacturing contract in China. "But we still need (US) equipment, materials, IP addresses and chip design software.It is unlikely that any of the world's chip makers .. is getting rid of American sellers soon. "
This chart from IC Insights clearly shows how flawed the market for chip design firms is.
SMIC, for example, is currently focusing on increasing production at the beginning of 14 nm, with plans to build a $ 10 billion smelter to focus on 14nm production for customers local. Assuming this facility is online and fully operational by 2022, China's production of integrated circuits will be about five years behind the rest of the world in terms of node leadership and seven to eight years in terms of volume production. However, according to reports, China 's overall self – sufficiency rate in semiconductor production would be only about 15%, due to the globalized and fragmented nature of the chain of custody. global supply.
Made in China 2025 plan demands 40% of Chinese Semiconductor Built in China by 2020 and 70% by 2025. According to experts, the currently achievable target is 20.5% by 2023. According to Nikkei, one of the obstacles China faces is that its own developers are more reluctant to rely on its local sites. technology produced or pay the higher prices associated with smaller-scale production from local fabs. This is another piece of this puzzle: companies do not perceive the Chinese product as an equivalent product, so they are less likely to buy it.
All of these preferences and positions are of importance, as long as China has a way to buy semiconductors made in the United States. If these roads closed completely, that would be another story. Having no other choice than to rely on its own know-how in semiconductor design, China would probably be late, probably for years. But such a decision from the United States would also attract considerable investment from the Chinese government, which would probably feel like they have no choice but to spend money to reproduce these pieces of the technological puzzle. that he could not buy elsewhere.
This is the central risk of any business, at least from the point of view of commercial enterprises. US companies that spend billions on developing new technologies do not want to give way to Chinese companies because of the cost of doing business. At the same time, however, they do not want to be totally excluded from these markets. Freezing the Chinese global semiconductor market by stifling access to American technology will not bring the Chinese back to vacuum tubes (although I would not mind seeing the first and only phone cell in the world powered by vacuum tubes). However, this could trigger a series of geopolitical events that have led other countries to agree to move away from US technology companies and sources if they see us as uncertain and unreliable partners.
Historically, it is extremely difficult to prevent the technology from wanting to acquire it. The United States could not stop other countries from inventing the atomic bomb after the Second World War. Firms with strong market conditions and near-monopolies in software distribution can benefit from these conditions, but even in these sclerotic markets, the change eventually occurs. Microsoft and Intel defined the PC sector from the late 1980s to the late 2000s, but now share space with various computer peripherals based on other operating systems and running different processors. architectures. Stuxnet – one of the most sophisticated malware programs ever created – has successfully delayed Iran's nuclear ambitions, without however preventing them alone.
Ironically, the impact of the current export control decisions by the United States is as unknown as the conditions governing other long-term bets made by semiconductor manufacturers. As Robert Palmer, former CEO of Digital Eleven, said: Designing microprocessors, it's like playing Russian roulette. You put a gun on your head, pull the trigger and discover four years later if you blew your brains.
It is always possible, of course, that the United States and China agree to resolve their trade disputes and that Chinese companies regain access to US technology. If that does not happen, however, we will be in Russian roulette for the long-term consequences. No country can afford to be completely cut off from the technology and semiconductor markets, which means that China would have little use other than to build its own knowledge base. and its own components, possibly as part of a broad international coalition with partners fearing to be treated. likewise by the United States.
There is no doubt that refusing to work with the Chinese would significantly slow down China's technological development – for a while. If that presents a long-term benefit to the United States – that is the question, is not it? Part of the argument is that the Chinese will recognize that they can not live without access to US semiconductor technology and that they will have enough incentive to play ball. But it is also possible that the Chinese, even if they are forced to return to the negotiating table for the moment, file their own long-term plans to move away completely from the technological dominance of the United States. . None of this means that the US declared objectives in the current trade war are less important, but they clearly show how problem-solving of this type can be complex.
Background image by SMIC.
The United States on Friday added five Chinese companies to a blacklist, limiting their access to US technology. The "list of entities" identifies entities for which there is a reasonable basis to believe, based on specific and explainable facts, that are involved, involved or have a significant risk of being or participating in activities contrary to US foreign national security policy ".
The companies in question are: Sugon, Higon, Chengdu Haiguang Integrated Circuit, Chengdu Haiguang Microelectronics Technology and Wuxi Jiangnan Institute of Computer Technology. One of them, Higon (also spelled Hygon), is a semiconductor joint venture between AMD and THATIC, in charge of selling x86 processors for the Chinese server market. THATIC itself consists of two separate joint ventures: Chengdu Haiguang microelectronics technology and Chengdu Haiguang integrated circuit design. If you look at the list above, these two companies are included.
In short, the US government seems to have just banned THATIC, the two companies that formed it, and the non-factory semiconductor company to design and sell x86 processors in China. According to the government's information document, all of these companies are owned in part by Sugon, which seems to be the main target company. The document of the Ministry of Commerce states It was decided that Sugon and the Wuxi Jiangnan Institute of Computer Technology were involved in activities:
United States … Sugon, the Institute of Computer Technology Wuxi Jiangnan and the National University of Defense Technology (NUDT) are the three main Chinese entities in the development of exascale high performance computing . Sugon has publicly acknowledged a variety of military end-uses and end-users of his high-performance computers.
This would seem to be a serious problem for AMD's efforts to expand its operations in China as part of the THATIC joint venture. We contacted AMD and received the following statement:
We are currently evaluating the addition of five new entities to the list of entities by the Bureau of Industry and Security. AMD will comply with the regulations governing this list, just as we have complied with US laws to date. We are reviewing the order details to determine the next steps for our joint ventures with THATIC in China.
Reports of earlier this month stated that AMD did not license Zen Architecture 2 to THATIC. The future involvement of AMD, its revenue forecast and exposure to it are currently not clear.
The move comes as President Trump prepares to meet Chinese President Xi Jinping in China next week and discusses how export laws can be used to radically change the competitive landscape. These determinations will change the balance of power between AMD and Intel in China, as well as the market available for Semiconductor in this nation the NYT reports that the United States is considering adding Hikvision to the list. Sugon is a leading exascale computer manufacturer in China, with 10 of the fastest Chinese supercomputers according to the TOP500. Sugon is a much smaller company than Huawei, but these specific lists will continue to impact the semiconductor industry given their exposure to the HPC market – a critical area in which AMD hopes to gain market share.
There was a time when computer capabilities were considered a national resource of the United States, and strict restrictions were imposed on the export of computer technologies. As Steve Jobs has already said: in 1999"The Power Mac G4 is so fast that it is classified as a supercomputer by the US government, and we are prohibited from exporting to 50 countries around the world." These restrictions have been considerably relaxed over the years. following, but some steps have been taken to tighten them since then. In 2015, for example, the Obama administration banned Intel, Nvidia and AMD from selling chips to the Chinese government. The Trump Administration's decision to further tighten the rules and lock down some subsidiaries is a welcome expansion of these restrictions, but not without precedent.
Once upon a time (2018), Intel and Apple were the best of friends contractually obliged to say nice things to each other from the public scenes as part of an agreement in which Apple would use the Intel 5G modem in its future products and where Intel would have an important customer to justify investing from aggressive way in 5G modems. The deal failed when Intel failed to meet certain deadlines or reached Apple's target targets. The details are uncertain. What we do know is that Apple has paid $ 4.5 billion to Qualcomm to settle the case and get a license for Qualcomm's 5G modem technology, and Intel has announced the closure of its modem and research and development activities.
Now there is to the rumor Apple might actually want to buy Intel's modem division in the first place, which would allow Chipzilla to no longer occupy the German design unit when it acquired Infineon. Intel, meanwhile, is potentially looking for a buyer. The company reportedly stated, "We hired external advisors to help us evaluate the strategic options of our 5G wireless business, and we have created value in both our wireless modem portfolio and our wireless portfolio. intellectual property. "Some Infineon engineers and executives are now Apple employees, making the deal more likely.
A few years ago, I wrote a pair of articles on the lack of success of Intel in the mobile market. I support both. I think they are resisting today as a result of discussions on some of Intel's missteps. But if I wrote the series again today, I could include the third part, the impact of Qualcomm's business practices on Intel's ability to find a market for its products.
Ars Technica has published an excellent deep diving in the anti-trust conclusion of Qualcomm last month that is well worth a read. The more than 200-page decision sets out a number of measures that Qualcomm has taken that, according to Justice Lucy Koh, were flagrant breaches of competition law. Qualcomm has structured the contracts to prevent competition from other chip manufacturers and customers who refuse to accept them may be exposed to a loss of access to the product.
There are some similarities between what Intel allegedly did to AMD in the early 2000s and what Qualcomm did with its own customers over the past two decades, including discounting practices that forced customers to sell Qualcomm equipment. 85% or 100% to receive them. This does not justify such practices and it is well known that Qualcomm dominated the LTE market, especially in the early years of the standard.
Apple would work on its own 5G modem even before separating from Intel. The idea that the company could buy the Intel Modem Division and make it the heart of its own networking efforts would hardly be a surprise. This would mean that Apple will finally have all the blocks needed to build a full internal processor (minus block input / output blocked, it may be licensed by some companies). With the fall of the iPhone, Apple with its own processor, graphics processor and modem would allow the company to generate a little more profit on each phone, helping to maintain margins and profits to a higher level when sales stagnate or fall.
As the launch date of AMD's next Zen 2 architecture is fast approaching, the company has raised the curtain and provided insight into the capabilities and improvements of its new model. These new chips include a number of enhancements and benefits to generate both higher cycle instructions (CPIs) and better overall power.
Let's start with some basics. The Ryzen 3000 family is powered by AMD's Zen 2 architecture except for its APU. APUs actually have a generation behind them – the 2000 Series APUs were actually built on the first generation of Ryzen, and the APUs of the 3000 series are based on the second generation Ryzen. The architectural improvements and other features we are going to talk about today do not apply to the Ryzen 3 3200G or the Ryzen 3 3400G.
During his Zen 2 presentation, AMD Corporate Fellow Mike Clark said his 7nm transition was in fact more successful than originally planned.
Some of you may remember rumors that AMD would use Ryzen 3000 processors with clocks much higher than the previous ones. According to AMD engineers, the company did not necessarily expect Zen 2 to reach higher frequencies. at all. This is the intrinsic problem of shrinkage of modern UC nodes. Smaller process requirements mean lower voltages, and lower voltages can negatively impact the absolute operating frequency. In this case, however, the 7 nm TSMC node and AMD's own engineering were able to create components that could reach slightly higher frequencies than the 12/14 nm chips.
It should be kept in mind that AMD did not expect improvements in the clock frequency of 7 nm, which should be taken into account when doing the same. Evaluation of the accuracy of rumors about massive clock jumps in the future.
An important change coming with Zen 2 has nothing to do with the current processor. AMD informed us during the event that new scheduler changes were being incorporated into Windows 10 Scheduler from Windows 10 1903 (May 2019 update). There are two new features: topology recognition and faster clocking of the clock. A faster acceleration of the clock reduces the time required by the processor to change state, thus improving performance and theoretically, inactive power by allowing the processor to move more quickly to different states. lower clock. Knowledge of the topology should help to keep local CCX data relevant and complete a CCX before loading another one.
These gains – a performance of + 15% in 1080p in Rocket League and a 6% improvement in the launch of PCMark 10 applications – are only result from the update of the Windows 10 scheduler and are separate from any additional gain resulting from improvements to the Zen 2 architecture. Taking advantage of these enhancements requires both an updated chipset driver and the upgrade. day Windows 10 1903.
This slide represents the microarchitectural synthesis of AMD. The chip integrates with the new TAGE branch predictor in addition to the BP perceptron it used in the past. The micro-op cache has been increased to 4K instructions, with double the total of L3 on board. (AMD is now referring to its L2 and L3 combinations as "AMD GameCache".) A new Address Generating Unit (AGU) is now associated with the entire side of the core, with full support of the comma floating 256-bit via AVX2.
The slideshow below presents our in-depth analysis of the specific architectural improvements of the third-generation Ryzen processor. Each slide can be clicked to open it in a new window.
According to AMD, these improvements give them an edge over Intel, both in terms of performance per watt and absolute wall power.
Cinebench is not the ideal solution for measuring energy consumption, but it is not a bad test either. The 3700X – which, in all fairness, is probably closer to the ideal position for architecture – is assumed to be 56% more efficient than the Core i7-9700K, while it only consumes 86 % of the power in absolute value.
Energy efficiency gains over the 2700X are even greater. AMD claims that the 3700X is 1.75 times more efficient in performance / watts than the 2700X, while consuming 70% power.
Although we obviously can not decide to launch before having material to test, AMD offers an aggressive and exciting product family. The TDPs have dropped dramatically. The CPI would have risen 1.15 times. Clock speeds have been increased. Planner improvements and doubled floating point capacity should provide their own robust improvements beyond this figure of 1.15. The width of the Infinity Fabric bus has been doubled to allow the use of PCIe Gen 4 bandwidth and a new memory divider at the DDR4-3733 level can reduce IF clocks without compromising the time of day. DRAM scale.
If you're a fan of AMD APUs, 7 nm also has exciting long-term implications for them. Although we do not know when we will see these parts, the company has clearly aggressively targeted a decline in power in all areas. This will clearly pay off when refreshing the APU family at 7 nm. According to one of our 7nm launch theories, AMD would focus on energy efficiency at least on some parts, and we see it perfectly, with a higher-performance 8-core processor in a 65W TDP and a 16-core processor in at 105W TDP.