DAVID FINCH: This is your EETimes Weekly Briefing. Today is Friday, April 19th, and among the top stories this week: Samsung is moving toward a 5-nanometer foundry process; Underwriters Lab is collaborating with Edge Case Research to draft a standard for autonomous systems. And we take a peek into Finland’s "Radio Valley" to learn about “Life after Nokia.”
DAVID FINCH：又到了EETimes全球联播时间。今天是4月19日，星期五。本周热门新闻故事有：三星正朝着5纳米晶圆工艺迈进；UL与Edge Case Research合作起草自动系统标准。然后我们将进入芬兰的“无线谷”，了解“后诺基亚时代的生活”。
Later in the show, we'll be joined by Phil Koopman, CTO at Edge Case Research and a professor at Carnegie Mellon University. Junko Yoshida, EE Times' chief international correspondent, asks Phil what makes autonomy standards-- whether developed at UL or ISO-- so hard to develop.
稍后，Edge Case Research的CTO和卡内基梅隆大学教授Phil Koopman将加入我们的谈话。EE Times首席国际记者吉田顺子（Junko Yoshida）询问Phil自动标准包括什么？是在UL还是ISO开发？以及开发的难度等问题。
We'll hear a report by Sally Ward-Foxton from our London bureau, who recently had an exclusive chat with Nigel Toon, CEO of GraphCore, the U.K.-based AI accelerator chip startup.
接下来让我们听听来自EETimes伦敦办事处的Sally Ward-Foxton的报道，她最近与英国的AI加速芯片初创公司GraphCore CEO Nigel Toon进行了独家采访。
And at the bottom of the program we will take a peek into Finland’s "Radio Valley" to learn about “Life after Nokia.”
All that to come, but first, Rick Merritt, EE Times’ Silicon Valley bureau chief, discusses Samsung’s announcement that it has completed its work on a 5-nanometer foundry process using extreme ultraviolet lithography. The announcement was a pre-emptive strike by Samsung. TSMC--the Korean giant’s largest rival-- is scheduled to give us an update on its 5-nanometer node next week. Rick explains what’s at stake in the upcoming finer-node battle.
RICK MERRITT: Samsung’s foundry group announced this week its 6-and 5nm process nodes are ready. And this is important because it's the latest data points from the leading edge of work on the traditional process of shrinking silicon that drives much of the electronics industry.
Samsung and TSMC are in some ways in a neck-and-neck race. They're both using the latest extreme ultraviolet lithography tools in production processes that offer something that's arguably can be called 7or 6 or 5nm process. And clearly, these capabilities are becoming more complex and costly to deliver, and that’s driving a change in how chip and system makers innovate.
So Samsung's an impressive chip maker and has deep experience making DRAMs and NAND flash and logic. It falls short as a foundry in that it doesn't have all the kind of IP blocks and partnerships and services that a TSMC can offer, or a size of a TSMC, which is three times as big a foundry compared to Samsung. And it lacks the blue-chip customer list that TSMC commands, given its status as a leading foundry.
Samsung knows where it falls short, and it's driving to be one of the leading-edge chip foundries. one of the last companies standing when Moore’s Law finally crashes into the atomic limits, perhaps around a 2nm node, let's say in 2024.
This is Rick Merritt in Silicon Valley for EE Times On Air.
DAVID FINCH: UL’s move to take a plunge into autonomy standards might have surprised some in the engineering community, especially those of us more seasoned.
Junko sat down with Phil Koopman, CTO of Edge Case Research and author of the UL 4600 draft. Phil initially approached EE Times to reveal what has been going behind closed doors at UL.
JUNKO YOSHIDA: Hi, Phil. Thank you for coming to the show.
So we learned from you this week that Underwriters Lab, UL, in close collaboration with Edge Case Research, is working on UL 4600 for autonomous systems. In parallel, we reported last month that ISO is working on a separate standard called SOTIF for ADAS and Autonomous vehicles.
Here’s my first question, Phil: What makes any autonomy standards-- whether done at UL or ISO-- difficult to develop?
PHIL KOOPMAN: Well, Junko, any autonomy standard has to grapple with the fact that this is a rapidly evolving technology area. The last thing you want to do is have a standard for each progress. ISO 21448 SOTIF is at its best when you have a set of requirements and you can apply more or less traditional safety engineering approaches. That means they concentrate on squeezing out as many of the unknowns as possible. And that way, what you're left with are unknowns that you can apply traditional safety to.
But as you move out of ADAS into full autonomy, the UL 4600 draft goes further in a couple directions. First, it encourages managing the risk of unknown unknowns based on continual field feedback. So you may ship with an unknown, but you're going to find out about it and fix it promptly to make sure the risk is minimized.
Second, it helps make sure the designers have thought of all the not-to-obvious implications of removing the human driver. A lot of 4600 is just a list of things to remember to think about, so design teams don't have to relearn lessons the hard way that other industries and other companies have learned. In other words, it helps you avoid putting vehicles at unnecessary risk.
JUNKO YOSHIDA: Now, here’s what I think gets interesting. Once it is not a human being but a machine that is driving an autonomous system, there are a lot of things system designers must think about in advance and prepare for safety.
In our interview, when you and I talked, you mentioned a couple of times, “Did you think of that?” Give us those “did you think of that?” examples please.
PHIL KOOPMAN: We found a lot of these, but I can give you a couple of examples to give you an idea.
First one: In a regular car, designers can take credit for the driver doing the right thing if something goes wrong. For example, if you lose your normal braking system, the driver is supposed to use the parking brake to stop. And by the way, this actually has happened to me, and I did use the parking brake to stop. If you put an autonomy system on a car platform like that, I certainly hope that the autonomy system knows how to do the same thing. What this means is that if you're combining an ISO 26262 approach with autonomy, all the credit that got taken for the human driver doing the right thing-- in other words, 26262 controllability-- now gets heaped on to the autonomy.
A second, perhaps more subtle example is that things in the supply chain, or even small maintenance mistakes, can go from being an annoyance to life-critical. As an example, let's say that someone in the supply chain waters down the windshield wiper fluid. They figure they're going to sell them summertime and it's no big deal if it doesn't have enough antifreeze in it. So you refill your car in the summer, and things are fine for a while. Or maybe you don't go down to the store and get fluid, you just put some tap water in. Either way, ends up the same place. But along comes winter, and some mud splashes on the windshield and the wiper fluid is frozen so you can't clean your windshield. Not a lot of fun in a human-driven car, but something like this has probably happened to a lot of us, and we figure it out. But if you're in an autonomous vehicle, and let's say you're the passenger asleep in the back seat, and all of a sudden all the cameras and the lidars are coated with mud, and that wiper fluid isn't there to clean things off. Now you've got a problem. You've got a vehicle going at speed that can't see.
Now I'm not saying that you have to solve it one way or another. Maybe there's some radar; maybe you have some other plan. That's great. But the question is, "Did you think of that?" Are you sure that you're going to be safe when that happens? Because it will eventually happen.
JUNKO YOSHIDA: UL 4600-- the draft you are working on-- is said to cover “validation of any machine learning-based functionality, and other autonomy functions used in life-critical applications.” Now here's my third question: Can you actually validate machine learning-based autonomy, and if so, how?
PHIL KOOPMAN: We've got an industry betting tens of billions of dollars that they can pull it off. So I certainly hope so! But to answer your question, Junko, the good news about the UL 4600 approach is that it gives companies a lot of flexibility. So long as the story they tell makes technical sense and they can say, here's why we're safe and here's why you should believe us, that's great. But this is a two-edged sword. To get all that flexibility, that means they have to decide how they want to argue they're safe.
So I expect that's going to entail cleverly limiting the operational design domain, and probably it will involve having a system that's pretty smart about knowing when it doesn't really know what's going on, and so it can be cautious.
It also talks about good safety hygiene and the types of things that are going to make it easier to explain why you're safe.
But beyond that, UL 4600 does not tell you how to build the system, it tells you that you have to explain why you're safe. Actually getting there? That's up to the designers.
JUNKO YOSHIDA: Phil, thank you so much for coming to the show. We really appreciate it.
DAVE FINCH: And now to our London bureau with Sally Ward-Foxton, who shares what she learned from GraphCore’s CEO. The outspoken executive now claims, “GPUs are effectively holding back innovation in developing futuristic neural networks in AI.” Here's Sally with more.
SALLY WARD-FOXTON: GraphCore are a British startup working on AI accelerator chips. They are an intriguing company, partly because they’re said to be the only semiconductor unicorn in the western world. At the last funding round, their company valuation was 1.7 billion dollars. Their investors include Dell, Bosch, BMW, Microsoft and Samsung, amongst others, and they’ve raised 300 million dollars so far. So obviously investors see something exciting in what GraphCore is doing.
The other reason GraphCore is so intriguing is their technology. They’ve been working with some of the leading minds in neural network research, and they’ve built a processing architecture specially for AI. It’s called Colossus, on account of it allows for building absolutely enormous processors. One GraphCore chip, one intelligence processing unit with the Colossus architecture, has 1200 separate cores, with each one running six program threads, so each chip can deliver 125 teraFLOPS of computing power. The numbers are absolutely staggering.
There are several important things I came away with after talking to GraphCore’s CEO, Nigel Toon.
First was the realization that neural networks are very diverse. Just as machine learning is applied to very varied types of tasks, neural networks for those tasks are obviously very different. When we’re talking about building a general purpose chip that can accelerate all different types of neural network, it’s much harder than building chips to accelerate specific AI applications. It’s this general purpose accelerator chip that GraphCore is working on.
Secondly, GraphCore is looking not just at the present, but towards the future of AI. As deep learning evolves, we’ll use techniques like reinforcement learning, where data is fed back through the network more than once, so the system can learn from experience. This is important when we want it to try and learn something about context, maybe if it was understanding spoken conversations, for example. As a result, we might end up with less separation between training chips and inference chips, because the inference side of things will be expected to learn as it gains more experience.
Finally was the idea that the incumbent technology for machine learning, GPUs, are effectively holding back innovation in developing these futuristic neural networks. Nigel Toon made the point that GPUs are great if you’re doing a relatively basic feed-forward convolutional neural networks, maybe looking for certain objects in photos. But for emerging techniques like recurrent neural networks and reinforcement learning, they don’t map well to GPUs. He said that whole areas of research are effectively being held back because there isn’t a good enough hardware platform available, and the fact that people are building application-specific chips and playing with FPGAs for machine learning research is a clear sign that GPUs aren’t up to the task. Hence the need for GraphCore’s IPU processor.
GraphCore, of course, like any startup company, they face several challenges. They’ve built a massive war chest, but they need to be focused on getting return on investment for their investors. Part of the problem will be growing quickly enough. They are currently at 200 people, and Nigel Toon says they’ll look to double in size this year. They have the cash, but finding that amount of qualified silicon engineers and software engineers in the space of a year will not be an easy task.
This is Sally Ward-Foxton reporting from London for EETimes.
DAVE FINCH: And finally, our London correspondent Nitin Dahad recently visited Oulu, Finland.
We often wonder what happens to a company town, anywhere in the world, after the anchor company closes its R&D facilities, shuts down its factories and lays off thousands of workers.
Here's Nitin's report on “Life after Nokia” in Oulu.
NITIN DAHAD: As a visitor to the city of Oulu in northern Finland, with its 200,000 population, three things defined it for me (apart from the heavy snow): the university, Nokia and saunas.
Yes, Finland is the land of saunas, with 2 million saunas for 5 million people, and it's an integral part of daily life, which I’ll come to later.
Oulu is a small city where a lot of business activity happens within something like a 20-mile radius.
So, when Nokia closed its mobile phone business around 2012, it was almost like an earthquake hit the city, as some 7,000 people lost their jobs. It was a real shock to the system, since the city was highly dependent on Nokia.
And even worse to come was the closure a couple of years later of Broadcom’s and Microsoft’s facilities.
I mentioned the University of Oulu. Its strength in radio technologies and microelectronics have been a key foundation for innovation, both in the Nokia era and the post-Nokia era.
I saw that with the work they are doing in both 5G and beyond (which people are now starting to call 6G). And it’s not just operating alone. It has a collaborative network of public and private partners.
This in my view led the foundation for the city’s regeneration.
I’ve been involved on and off in economic development for many years, and in the late 1990s, governments around the world started looking at the success of Silicon Valley in California and trying to figure out how they could emulate the same formula and create the next Silicon Valley in their own countries.
It’s common wisdom that the key ingredients for the melting pot are knowledge, talent and money.
In Oulu, the university provided the knowledge foundation. And the talent came from the experience gained at Nokia over many years.
The next bit was the money, and that’s where there appeared to be (at least this is what city officials tell us), a coordinated program from Nokia and local government to fund startups who had ideas.
The whole region effectively became a radio incubator, with entire teams who were laid off starting their own companies-- such as Testilabs, who bought out lab equipment from Nokia and started offering radio test services to the industry as an independent service. Others include active noise canceling earplugs developer Quieton, and Haltian, a hardware design engineering firm focused on IoT.
You hear many such stories as you wander through the corridors of some of the office and lab units in Oulu, which were once themselves home to Nokia’s engineers.
I mentioned three ingredients for successful economic development: knowledge, talent and money. But if it was so easy, every city in the world would be able to do it. There’s more to it: you need the networks and the culture, as well as customers.
The ex Nokia-staffers had strong networks and a culture that enabled them to collaborate when outside Nokia. This is probably where the saunas come in. Saunas are an integral part of the way of daily life of Finnish people. KNL Networks, for example, was established after the founders had the idea while sitting in a sauna.
Saunas aside, I think it creates strong bonds of mutual trust among people, and this is what ultimately is essential for doing business.
So when outside Nokia, people were able to call upon their networks to do the design, the validation and testing, and the manufacturing. And be sure that services and products would be delivered according to market needs and when they needed it.
Many of the startups coming out of Nokia were able to reap the rewards of these networks quickly, some getting contracts with key blue-chip customers around the world within a year. This created a rich ecosystem of companies in the entire supply chain in Oulu, creating what I like to coin a "Radio Valley."
The regeneration and development model I mentioned is repeatable, and it's been done with varying degrees of success. Israel’s Tel Aviv and Haifa are probably the most well-established and successful examples. Then you have Bangalore, London, and in my travels, I’ve seen similar emerging clusters in places like Cairo and Alexandria in Egypt, Sao Paulo in Brazil, and many others around the world.
This is Nitin Dahad, EE Times’ European Correspondent, reporting from Oulu, Finland.
DAVE FINCH: That was Nitin Dahahd, and this has been your weekly briefing from EETimes and the AspenCore Global Service. You can read all these stories and more at EETimes.com. Thanks for listening.