DAVID FINCH: This is your EETimes Weekly Briefing. Today is Friday, April 12th, and among the top stories this week - Qualcomm’s new data-center AI inference accelerator chip, the latest deep learning developments unveiled at a Stanford University gathering, and a new development that throws a monkey wrench into the debate about V2X – vehicle to everything-- in Europe.
(DAVID FINCH：又到了EETimes全球联播时间。今天是4月12日，星期五，本周头条新闻有：高通新的数据中心AI推理加速芯片、斯坦福大学会议揭开最新深度学习进展，还有欧盟的一项最新决议终止了V2X(车辆到一切)的争论。)
Later in the show, Echo Zhao, China Chief Analyst of EE Times China,talks about an industry discussion that her team recently organized in Shenzhen under the topic, "Countdown: How Far is China from 40% IC Self-Sufficiency?"
(接下来，电子工程专辑中国首席分析师赵娟谈到最近在深圳(译者注：应该是上海，中国IC领袖峰会)举办的一个行业论坛，主题是“倒计时：中国离40％的IC自给率还有多远？”)
And Steve Taranovich tells us about AspenCore Media’s latest Special Project on “reusable rockets.”
(然后Steve Taranovich给我们介绍AspenCore媒体集团最新的专题项目 --可重复利用的火箭。)
All of that to come, but first, Dylan McGrath, executive editor of EETimes, this week attended Qualcomm’s event in San Francisco. We asked him to relay what Qualcomm said at its announcement and, perhaps more important, the details Qualcomm didn’t mention in its announcement. Here's Dylan with more.
这一切都会逐一为你诚献，但让我们首先听听来自EE Times执行主编Dylan McGrath的消息，他本周参加了高通在旧金山举办的活动。我们问他高通公司在新闻发布中说了什么，或许更为重要的是，高通未提及的细节。请听Dylan的报道。
DYLAN MCGRATH: The title of Qualcomm's event was "Qualcomm AI Day 2019," and the headliner was the introduction of Qualcomm's first AI inference accelerator, and it's specifically for the data center.
So a lot of interesting things there. There were not a large number of details given. About the only hard number specifics that Qualcomm shared was that this is a 7-nanometer chip, it has a very impressive, more than 350 trillion operations per second peak AI performance, and it also promises more than 10x the performance per watt of anything that's deployed today.
All of those things sound very enticing, but the chip's not going to be sampling until the second half of next year. It's likely not going to be in production until late next year. And it's really hard to say how the market's going to evolve between now and then.
We are aware that there are more than 30 companies that are currently working on purpose-built AI inference accelerators. And Qualcomm,really no surprise, threw its hat in the ring. And the company does have its heritage as a mobile chip vendor, and it really tried to emphasize at its event that this will be a power-efficient alternative to the Nvidia chips and some of the other things that are out there and coming out.
But again, without more hard numbers it's really hard for analysts to say how successful this chip will be. I think, at the end of the day, if Qualcomm can deliver on the type of power efficiency that it's promising, there will be a market for this chip. Probably a pretty substantial market. If it's unable to deliver on that type of power efficiency, it's hard to see that they'll be much market for this chip at all.
This is going to be a huge market. Projections for AI inferencing silicon in data centers go as high as $17 billion by 2025. Qualcomm, of course,wants a piece of that, as does everyone else, which is why they're such intensecompetition.
DAVID FINCH: That was Dylan McGrath reporting from San Francisco.
Rick Merritt, EETimes’ Silicon Valley bureau chief, recently attended a gathering called SysML held at Stanford University in Palo Alto.Here’s what he filed from the field.
RICK MERRITT: I recently attended the second annual SysML at Stanford. It’s a really good conference at the nexus of deep learning and running deep learning jobs on really large scale data centers.
So this was started a couple years ago by a group of the big data centers like Google and Facebook and Microsoft and Amazon, along with a lot of top drawer academic researchers and a sprinkling of semiconductor companies involved in this area like Intel and Nvidia. So it's really interesting in part because the big data center folks are doing a good job of getting in there and very frankly sharing their learnings, trying to make deep learning work on these really big networks of computers. So there's a lot to sort out there, a lot to learn, and a lot that people are sharing.
My favorite paper from the event this year was from researchers at the University of Texas at Austin. And typical of what I'm hearing and other papers there, they not only shared some techniques that are kind of at the application level-- in this case what they called mini-batch serialization, away of packaging operations together for efficiency-- but they also talked about a new architecture for doing it, which they happen to call WaveCore,still an academic microprocessor architecture, still a simulation-only chip,but one they have some thoughts about possibly trying to commercialize. The cool thing about that combination was that they were able to handle training performance levels at or in excess of today’s top-end Nvidia V100 chips can do,but needed less memory to do them. They actually could operate them on mobile DRAM.
So good work that's being shown, and I highly recommend a trip to SysML.
This is Rick Merritt in Silicon Valley for EETimes.
DAVID FINCH: AspenCore’s co-global editor in chief Junko Yoshida wrote about V2X, particularly a recent vote by European Parliament’s transport committee. She explains what’s at stake and why it matters.
JUNKO YOSHIDA: Communication between vehicles to vehicles, or vehicles to infrastructure, generally known as V2X,has long been viewed as an important safety layer, alerting cars to up coming obstacles even before they are visible.
The industry has developed communication protocols and wireless technology based on IEEE 802.11P, known as Dedicated Short-Range Communications, or simply DSRC. Specifically designed for vehicular communication, DSRC is sort of like WiFi. It’s royalty free and it’s gone through rigorous testing for ten years. In my opinion, what's not to like?
All that groundwork, however, hasn’t stopped the cellular industry from trying to hijack the V2X agenda, claiming ownership and hiring an army of lobbyists. The goal is to sway automakers, politicians and regulators into believing they’re much better off with cellular-based V2X. The future, they say, is not DSRC, but 5G.
Well, never mind that 5G is not ready yet.
A big surprise this week was that a transport committee of EU lawmakers Monday rejected a pro-DSRC draft proposed by the European Commission.
The cellular industry and some European governments are reportedly opposed to the current EC plan, because it requires a V2X system that supports“backward compatibility” with DSRC.
更多...
——听全球专业科技资讯，只在EETimes
用户评论