BRIAN SANTO: And I’m Brian Santo, EE Times Editor in Chief, and this is your Briefing for the week ending November 8th.
BRIAN SANTO: 我是EE Times主编Brian Santo，以下带来截至11月8日的本周简报。
AI companies do seem to be hogging the news lately, which makes the following interesting. One high-profile startup created the impression it was about to make some big news – but then didn’t. Startup Groq was a no-show at a conference it was co-sponsoring. Here at EE Times we've been trying to recall if there is a precedent for that. We stumped ourselves – and some of us have been involved with the high-technology since the 1980s.
最近AI公司似乎都在抢着当新闻热点，这使得接下来要讲的事情变得有趣。 一家备受瞩目的初创公司给人感觉像是要发布些重大新闻——最后却并没有。 这家名为Groq的初创公司没有出席由其共同赞助的会议。 我们EE Times一直在努力回忆，Groq这样做是否出于某种原因。这倒是把我们难住了——尽管自20世纪90年代以来，我们的部分同事就参与到高科技中。
Sally Ward-Foxton finally got a chance to talk to Groq recently. International editor Junko Yoshida talked to Sally about what the company said.
最近，Sally Ward-Foxton终于有机会与Groq进行了沟通。国际编辑Junko Yoshida将与Sally聊聊这家公司。
JUNKO YOSHIDA: Hi, Sally! How are you?
JUNKO YOSHIDA: 嗨，Sally！最近怎么样？
SALLY WARD-FOXTON: Hi, Junko! Good, thanks.
SALLY WARD-FOXTON: 嗨，Junko！很不错，谢谢关心。
JUNKO YOSHIDA: This interview actually I was really looking forward to. Because here's the thing: You and I or the EE Times write a lot of stories about AI. And especially AI startups, right? So when one of the most secretive AI accelerator startups in Silicon Valley doesn't show up at a conference for which the companies was one of the major sponsors, of course everybody starts wondering what's going on, right? So especially when that AI startup is Groq, which is, you know, backed up by big investment and it's got this well-known Google pedigree, everybody is sort of envious and also jealous. And everybody wonders what's going on. So tell me. You had this big scoop. You recently talked to Groq's top management team, right? What was their excuse? Or what was their... How did they explain for their no-show at the AI Hardware Summit in September?
JUNKO YOSHIDA:我真的很期待你这次采访。因为事实上：你和我，或是EE Times，都撰写了许多有关AI的故事，特别是AI初创公司，对吗？因此，当硅谷最神秘的AI加速器初创公司之一，没有出席由其担任主要赞助商之一的会议时，当然人人都会开始好奇，想知道发生了什么，对吗？特别是这家AI初创公司是Groq。Groq名气很大，而且得到了大笔投资的支持，还有了Google的“血统”，所有人都感到羡慕和嫉妒。每个人都想知道发生了什么。所以和我讲讲你这个特大独家新闻。你最近和Groq的高管团队进行了交谈，对吧？他们的借口是什么? 或者说他们是怎么，他们如何解释没有参加9月AI硬件峰会的原因？
SALLY WARD-FOXTON: Yes. So certainly from the outside it looked like they had chosen to come out of stealth mode and then basically chickened out. Also, between us colleagues, we speculated at the time about whether they were about to be acquired, but no, that's not what happened. Finally I got to speak to Groq to find out what really happened. I spoke to Jonathan Ross, Groq's CEO. He said they were dealing with a company, with a customer issue that took priority, which meant they didn't have time to finish the demo they were working on. So rather than just present the company with no demo, they decided to withdraw completely.
SALLY WARD-FOXTON: 好的。可以肯定的是，从外部视角看来，Groq似乎已经选择走出“隐秘模式”，然后就要出头了。另外，我们同事之间当时也在猜测，是否有公司要收购他们，然而并不是，事实上也确实不是我们猜想的那样。终于我找到机会与Groq聊聊，了解事情真实情况。我采访到Groq的CEO Jonathan Ross。Jonathan说，Groq正在处理和另家公司的问题，客户问题是放在首位的，这就意味着他们没有时间去完成展示，没能给大家展现目前的研究内容。所以他们决定完全退出会议，而不是单单不做展示。
JUNKO YOSHIDA: Right. Did you believe him?
JUNKO YOSHIDA: 好吧，你相信他说的话吗？
SALLY WARD-FOXTON: Well, I have no reason not to. I feel like if they were going to make something up, they could have made up something better. But it does seem a little bit odd. I mean, why not send somebody to present the company's slides, the company's pitch deck, and tell us something about the company. Because they were in stealth mode up to now. They would have had plenty to say. I mean, rather than just withdraw completely at such a late stage, which was a PR disaster... I mean it was quite embarrassing, given that they were a sponsor at the conference and everything, when they didn't show up. So yeah. A very strange decision.
SALLY WARD-FOXTON: 嗯，我没有理由不相信。我觉得如果他们要找借口的话，完全可以编个更好的理由。不过这确实有些奇怪。我的意思是，为什么不派个人去展示公司状况，并介绍下有关公司的信息。因为到目前为止，Groq都还处于“隐秘模式”。他们本可以有很多可介绍的内容。我是说，也远比在这么迟的阶段完全退出会议要好吧，这简直是一场公关灾难。我觉得这挺尴尬的，因为他们是这次会议的赞助商，他们赞助了会议，结果自己却没出席。是的，真是一个非常奇怪的决定。
JUNKO YOSHIDA: Right. So you're saying that Groq has never shared any architectural diagrams, or they have never shared any of the concept of the architecture before?
JUNKO YOSHIDA: 对。你有提到说Groq从未分享过他们的任何架构图，之前也从未公开过任何架构概念？
SALLY WARD-FOXTON: So up to now they haven't shared anything. The website was just one page with not really much on it. And they're still, even for a company that are coming out of stealth mode, they still are very secretive. Even the slides that they showed me during our meeting, they wouldn't let me have a copy of. There was a product photograph even they wouldn't let me have a copy of. So yeah, they're very secretive even still.
SALLY WARD-FOXTON: 到目前为止，他们还没有分享过任何东西。他们网站也只有一页，没有太多内容。 即使是对于一家正脱离隐秘模式的公司而言，它们也仍然是非常神秘的。他们在会议期间向我展示的幻灯片，也不会同意我拷贝副本。里面有一张产品的照片，都不允许拷贝给我。是的，他们仍然很神秘。
JUNKO YOSHIDA: So let's get back to the interview you had with Groq's management team. I think that involved not only the CEO but also a chief architect as well as the VP of Engineering, is that it?
SALLY WARD-FOXTON: That is right, yeah. I spoke to Dennis, Abts is the chief architect at Groq. I also spoke to Michelle Tomasco, the VP Engineering; and the COO, Adrian Mendes.
SALLY WARD-FOXTON: 是的，我和Dennis Abts进行了交谈，他是Groq的首席架构师。我还采访了工程副总裁Michelle Tomasco，以及首席运营官Adrian Mendes。
JUNKO YOSHIDA: Okay. All right. So how did they explain? How is Groq's approach to AI or AI hardware essentially different from a lot of its competitors?
JUNKO YOSHIDA: 好的，那么他们如何解释这件事呢？Groq的AI或AI硬件接入方面与其众多竞争对手有何本质区别？
SALLY WARD-FOXTON: So they didn't suffer to find hardware. They developed this compiler almost before they developed the chip, and the compiler takes on some of the tasks which would normally be done in hardware. So you basically feed in tensor flow, and then the compiler does all the execution planning and orchestrates the data flow and all the timings within the chip. Because such a lot of it is done in software, it means you can get rid of a lot of the hardware control functions in the chip to deal with things like synchronization. You don't need speculative execution, protection. It makes things a lot simpler.
SALLY WARD-FOXTON: 他们在找到合适硬件上没有问题。Groq差不多在开发芯片之前就开发了编译器，且编译器承担了通常需要在硬件中完成的某些任务。所以你基本上是输入tensor flow，然后编译器就会执行所有制定计划，并编排数据流和芯片内的所有时序。由于很多工作都是通过软件完成，这意味着你可以摆脱芯片中的许多硬件控制功能，来处理诸如同步之类的事情。你不需要推测性执行和保护。它使事情变得简单得多。
The advantages of doing it is this way is that everything becomes completely deterministic, completely predictable. There's no long-tail latencies. You know exactly what the latency will be and how much power you're going to need right from the start, right from compile time. And this predictability is very attractive to some applications like data centers and autonomous vehicles.
JUNKO YOSHIDA: Right. It is sort of like turning the whole approach to AI on its head in a way that you do software development first, but you focus a lot on the building of the software compiler. And then going back to the AI hardware. Is that it?
JUNKO YOSHIDA: 对，这有点像是将整个AI路径颠倒过来，首先要进行软件开发，但是你将重点放在了软件编译器的构建上。然后再回到AI硬件，对吗？
SALLY WARD-FOXTON: That's it. You can almost do all the software and use the compiler before you've got the silicon, so you can almost get a head start that way.
SALLY WARD-FOXTON: 就是这样。在拥有芯片之前，你几乎可以完成所有软件并使用编译器，因此你差不多可以利用这种方式抢先一步。
JUNKO YOSHIDA: That's actually kind of new, in my opinion. Because a lot of times when we talk to chip companies they say, they talk about chips in great detail and hardware. And then at the end of the presentation they usually say, Oh by the way, we've got this compiler.
JUNKO YOSHIDA: 我认为这实际上是一种新事物。因为很多时候我们与芯片公司交谈时，他们会谈到芯片的细节和硬件。 然后在演示产品的结尾，他们通常会说：“哦，顺便说一下，我们有这个编译器。”
SALLY WARD-FOXTON: So this is also about the compiler. The story is both about the chip and about the compiler. Certainly they think it's very new and novel, and as we find out more about the company we'll find out how new and novel it is. Certainly there are other companies in this space such as SambaNova that said they're doing software-defined hardware as well. Although we don't know anything about SambaNova really. They're still in stealth. So yeah, we'll find out as more details emerge how new and novel this is.
SALLY WARD-FOXTON: 所以这与编译器也有关系。这个故事既涉及到芯片也涉及到编译器。当然，Groq认为这是非常新颖的概念，随着对他们公司的了解更深入，我们将会发现这是多么新颖的想法。当然，在这个领域中还有其他一些公司，例如SambaNova，该公司也表示正在做由软件定义的硬件。 虽然我们对SambaNova一无所知，这家公司也是处于“孵化状态”。不过是的，随着更多细节的出现，我们将发现这种概念是多么新颖和创新。
JUNKO YOSHIDA: All right, so you just mentioned software-defined hardware. I think Groq also... the basic concept as you explained in your story is the software-defined hardware. You know, that sounds a lot like FPGA in a layman's point of view. But it looks like the Groq CEO is adamant that is not an FPGA. How did he explain that to you?
JUNKO YOSHIDA: 好的，你刚才提到了由软件定义的硬件。我认为Groq也...你在故事中解释的基本概念是软件定义的硬件。从外行的角度来看，这听起来很像FPGA。但是，似乎Groq CEO坚决认为，这不是FPGA。他是如何向你解释的？
SALLY WARD-FOXTON: The phrase, "software-defined hardware," combined with all this talk of determinism, predictability, really makes you think of FPGAs. But Jonathan Ross said to me, he said explicitly, "This not an FPGA." It's not an FPGA. It's not look-up tables or anything. And you can reconfigure the Groq chip every cycle, he said. I think it's a similar idea in terms of knowing what the chip is doing exactly all the time and reconfiguring it with software. But no, it's definitely not an FPGA.
SALLY WARD-FOXTON: “软件定义的硬件”一词与所有关于可确定性、可预测性的讨论结合在一起，确实让你联想到FPGA。 但是Jonathan Ross很明确地对我说：“这不是FPGA。” 它不是FPGA。 它不是查询表或其他任何东西。他说，你可以在每个周期重新配置Groq芯片。我认为这是一个类似于FPGA的想法，因为它始终知道芯片在做什么，并用软件对其进行重新配置。 但绝对不等同于FPGA。
JUNKO YOSHIDA: So where does Groq go from here? Can you tell us some of the roadmap that they shared with you?
JUNKO YOSHIDA: 那么Groq是怎么打算的？你能和我们讲讲他们分享给你的一些规划吗？
SALLY WARD-FOXTON: Well, I would if they had shared a roadmap with me. All we know at the moment is they've got the silicon back, which is now going into production. They say they're getting traction with customers already. They've apparently been shipping their chip on a PCIe board since August. So hopefully they can start to get some design wins. They're certainly... One of the things about the chip is that it can handle both inference and training, although they're very much focused on the inference market right now. So moving forward, like in a future generation of products, I guess they could start to tackle training. I think we'd all love to see that demo, as well, that famous demo that was made. Maybe they'll do some more of that and we'll finally get to see it.
SALLY WARD-FOXTON: 好吧，如果他们把规划共享给我，我会告诉大家。 目前，我们所了解到的是，他们已经拿到了硅片，目前已投入生产。Groq说他们已经吸引到客户。自八月以来，他们显然一直在用PCIe板运送芯片。希望他们能够赢得一些设计大奖。他们确实…关于该芯片的一件事是，它既可以处理AI推理又可以进行AI训练，尽管Groq现在非常专注于AI推理市场。往前看，正如看待下一代产品，我想他们可以开始着手AI训练功能了。我想我们都希望看到真实的产品展示，以及那个著名的DEMO是如何制作的。也许他们还会做得更多，我们最终将能看到成果。
JUNKO YOSHIDA: All right then. I guess it's one of those stories: Stay tuned! Right?
JUNKO YOSHIDA: 好的，我想这会是系列故事中的一个：敬请期待！ 对吗？
SALLY WARD-FOXTON: Yes. Stay tuned for more.
SALLY WARD-FOXTON: 是的，敬请期待更多内容。
JUNKO YOSHIDA: Okay. Thanks! Bye.
JUNKO YOSHIDA: 好的，谢谢！再见.
SALLY WARD-FOXTON: Bye.
SALLY WARD-FOXTON: 再见.
BRIAN SANTO: Sally has technical details in her story on eetimes.com.
BRIAN SANTO: Sally的这篇文章发布在eetimes.com，文中有讲到技术细节。