Grok Bots and How CPUs are used in Agentic AI

那你看看这个。

So check this out.

这周其实发布了Grok Bot，这是一个智能体AI平台。

So this week was actually the release of Grok Bot, which is an agentic AI platform.

你可以把它想象成OpenClaw，但用起来特别简单，这样每个人都能随时运行智能体，即使你的电脑是关着的。

So you think about it like an OpenClaw, but like extremely easy to use so that like everybody can run agents all the time right now even if your computer is closed.

我觉得这是一个非常重要的进展，倒不是因为Grok Bot这个产品本身，而是它背后所代表的计算底层，特别是对CPU的意义。

So this is a very important development, I think, because it's not so much for what the Grok Bot is as a product, but what it means from an underlying substrate of compute and specifically what it means for CPUs.

听众朋友们，大家好。

Hello, listeners.

欢迎收听新一期的Semi Doped节目。

Welcome to another Semi Doped episode.

我是来自Chipstrat的Austin，这位是来自Vic's Newsletters的Vic。

I'm Austin from Chipstrat, and this is Vic from Vic's Newsletters.

如果你是在YouTube之类的地方看这个视频，你会注意到Vic现在在不同的地方。

And if you're watching this on YouTube or something, you'll notice Vic is in a different location.

维克，告诉我们你在哪里？

So Vic, tell us where are you?

然后我还看到你在用那个Grok Bot，我在X上看到的，但我还没试过。

And then I also saw that you are trying this Grok Bot, which I saw on X, but I have not tried.

说实话，我对它一无所知。

I honestly don't even know anything about it.

所以我想听听你在哪里，然后跟我聊聊Grok Bot，因为我知道你一直在玩它。

So I wanna hear where are you and then tell me about Grok Bot because I know you've been playing with it.

太好了。

Awesome.

我现在在湾区。

I am in the Bay Area.

我来这里是为了下周的Hot Chips，非常激动。

I came here for like hot ships next week, which is very exciting.

很多人都会来参加这个会议，但我决定提前一周到，因为我之前其实在台北参加OpenCompute Project的亚太区会议。

A lot of people are coming down for this conference, but I decided to show up a week early because I was actually in Taipei attending the OpenCompute Project APAC, like the Asia Pacific region version of this conference.

我当时就想，提前来这边确实挺有道理的，因为在湾区要见的人太多了。

And I was like, yeah, think it makes just sense to like come here a little bit early because I've got so many people to meet in the Bay Area.

这几天真的太棒了。

It's been amazing.

我遇到了很多很棒的人，也发生了很多很棒的事，这些以后都会慢慢聊到的，等时机合适的时候我们会在播客上好好聊聊。

I've met so many amazing people and so many amazing things have happened that we should eventually it'll all come out at some point and we'll discuss it on the podcast or whatever when the context is right.

总之，这周过得特别精彩，我现在就在湾区。

But yeah, it's been an amazing week and so I am in the Bay Area.

我带了小相机设备，但没带我的Semi Doped牌子，这点挺遗憾的。

I got my little camera rig with me but I don't have my Semi Doped sign which I'm very bummed about.

所以下次我得带着我的霓虹灯牌出门。

So I'm gonna travel with my neon sign next time.

说到Grok Bot。

So like Grok Bot.

Grok Bot挺有意思的，因为它是一个智能体AI软件。

Grok Bot is an interesting thing because this is an agentic AI software.

怎么说呢，我也不太清楚。

Like, I don't know.

他们管这个叫智能体框架，就是这么个术语。

They call it an agentic harness is the term for it.

就是一个软件，你直接从网站上下载下来，然后安装。

It's a piece of software you just download off the website and you install it.

装好之后有个界面，基本上所有功能都是开箱即用的，对吧？

And you've got an interface where it pretty much everything works out of the box, right?

所以你要做的就是，打开它，然后它会问，嘿，你想让它做什么？

So what you do is the first thing that comes up and you go, Oh, hey, here's the, you know, what do you want it to do?

然后我就说，嘿，我想让你当首席智能体，你是我所有智能体的头儿，你来跟其他智能体沟通。

And I said, Hey, I want you to be a chief agent and then you're the chief of my agents and we'll you're gonna talk to all the other agents.

然后它就问，哦，你还想要哪些智能体？

And then it's like, oh, what other agents do you want?

我就说，哦，我想要一个能做这个的智能体。

I'm like, oh, I want an agent that does this.

我想要一个助手代理。

I want an assistant agent.

我还想要一个，嗯，新闻监控代理之类的。

I want a I don't know, a news monitoring agent, something like that.

然后它就会自动把所有东西都配置好，对吧？

And it just spins it all up automatically, right?

挺高级的。

It's fancy.

真的很不错。

It's really nice.

所以我觉得这会为很多人开始使用代理打开大门，因为之前OpenClaw和Hermes那套方法有点难搞。

So, I think this will open up the doors to a lot of people starting to use agents now because the whole OpenClaw approach and the Hermes approach was like little bit difficult.

你得安装它，还得进命令行，设置API密钥，还得配置各种东西。

Like you got to install it and you got to go into the command line and you've got to set up an API key and you've got to like set up all of this stuff.

对大家来说确实有点复杂，但Grok Bot把所有事情都搞定了。

It is a little bit involved for everybody to use, but Grok Bot takes care of everything.

不错，太棒了。

Nice, amazing.

好的，那像 OpenClaw 和 Hermes，不管怎么发音，它们都是在你自己的电脑上本地运行的吧？

Okay, so like OpenClaw and Hermes, Hermes, however it's pronounced, those ran locally on like your own machine, right?

那这个 Grok Bot，它有多少是在本地机器上运行的，又有多少是在云端呢？

And this Grok Bot, this must, how much of it runs locally on the machine versus in the cloud?

你说你下载了什么东西。

You said you downloaded something.

对，这只是一个软件，它本质上是在和云端的电脑通信。

Yeah, this is just a piece of software that essentially communicates with the computer in the cloud.

这就是它好的地方，对吧？

This is the nice part of it, right?

我真正喜欢它的一点是，它实际上是在某个服务器上运行你自己的专用小电脑。

Which is what I really like about this is it is actually running your own little dedicated computer in some server somewhere.

所以你有一个小虚拟机在某个地方运行，所有东西都在那个盒子里。

So you have like a little virtual machine that's running somewhere and everything is within that box.

你所有的工具都装在那里。

All your tools are installed there.

你所有的登录信息也都存在那里。

All your sign ins are installed there.

你不需要那些MCP连接，也不需要那些插件，这个插件那个插件的。

Like you don't need all these MCP connections and these plugins and those plugins.

你就把它想象成你自己的电脑，只不过是在云端。

It's just think about it as your computer, but somewhere in the cloud.

它就是一个虚拟机，对吧？

It's a VM, it's a virtual machine, right?

它的好处在于，所有安全措施和那些让代理在沙盒环境里运行的关键东西，全都包含在里面。

The nice thing about this is that all of the security practices and everything that matters for these agents to run-in its sandbox environment is all contained within there.

所以如果你的代理登录了某个工具，那它就能轻松地在所有代理之间共享，因为这就是你拥有的一台本地电脑。

So if your agent has like signed in into one tool, it is available easily across all your agents and everything because it's a local computer you have.

这就是它和OpenClaw、Hermes的区别。

That's the difference compared to like OpenClaw and Hermes.

不错，太棒了。

Nice, amazing.

听起来确实很方便。

That does sound so convenient.

那么，你觉得这比在本地机器上运行一堆需要自己管理的 Cloud Code 代理要好，是吗？

So in this, would you say it's better than just like having a bunch of Cloud Code agents running locally on your machine that you have to manage?

这其实是有得有失的，对吧？

So it's a give and take, right?

当你拥有自己的 Cloud Code 代理时，会有很多控制权。

There's a lot of control when you have your own Cloud Code agents.

你可以非常细致地指定很多东西，比如用哪个模型来做什么。

You can specify a whole lot of stuff to the finest detail, like which model you should use for doing what.

而且你可以优化 token，比如可以说，嘿，对于这些查询，你可以这样做，所以你可以通过 Cloud Code 界面把 OpenRouter 设置为模型提供商，然后在 OpenRouter 里你可以说，好的，用自动路由或者用这个模型层级，这样这些请求就被路由到这里，那些请求就被路由到那里。

And you can token optimize and be like saying, Hey, for these queries, can do, you know, so you can set up open router as a model provider through the Cloud Code interface and you can in open router you can say, okay, use the auto router or use this hierarchy of models such that these requests are routed here and those requests are routed there.

而这样做就把所有这些都省掉了。

This does this takes all that out of the equation.

就像你根本不知道发生了什么。

Like you just you don't know what's happening.

你只有令牌，然后你就用这些令牌。

You just have tokens and you use those tokens.

不错。

Nice.

就这样。

That's it.

你不知道用了什么。

You don't know what's being used.

你知道，我觉得简单按钮才是未来的趋势。

You know, I think the easy button is how it's gonna go.

智能体会走向大众，他们不需要去买Mac Mini。

Agents will go to the masses where they don't need to go buy a Mac Mini.

他们不需要安装任何东西，也不需要知道OpenRouter是什么，或者它可能现在要被Stripe收购了。

They don't have to install something, they don't have to know what OpenRouter is, or which maybe is getting acquired by Stripe now.

但他们不需要配置，也不需要设置最大令牌数，或者去搞什么令牌效率、最小令牌数。

But they don't have to configure and token max or get token efficient, token min.

最小。

Min.

他们只需要按一下那个简单的按钮就行了。

They they can can just hit the easy button.

那这是不是意味着我应该买一堆CPU股票？

So then does that mean I should buy a bunch of CPU stocks?

如果Grok要启动所有

If Grok is gonna spin up all

这些虚拟机？

these VMs?

我们很快就会说到“这不是财务建议”的部分。

We'll get to the not financial advice part really quickly.

但你提到了Mac Mini。

But you mentioned Mac Mini.

我其实就想提一下这个。

I just wanted to like touch upon that.

就像之前智能体刚推出的时候，其实也没多久，也就是去年十二月我们才听说这个东西。

Like there was this when agents were rolled out not so long ago, it was like only December last year that we even heard of this thing.

当时大家抢计算资源抢得特别凶，连英特尔都没预料到CPU的需求会这么大，对吧？

It was such a rush to get compute that even Intel was caught unawares of the demand for CPUs, right?

他们在二月份的财报电话会议上说，哎呀，我们CPU需求特别大。

In their earnings call in February, they said, oh, we have a lot of CPU demand.

我们完全没想到会这样。

We had no idea it was coming.

所以问题就是，为什么会发生这一切？

So the question is why did all that happen?

如果当时大家观察得够仔细，其实能看出点端倪，因为智能体一出来，所有人都跑去买Mac Mini了。

If people observed keenly enough at the time, you could kind of tell why what's happening because the moment agents showed up, everybody ran and bought a Mac Mini.

那他们到底在买什么呢？

So what exactly were they buying?

他们买Mac Mini是为了内存吗？

Were they buying a Mac Mini for its memory?

是为了里面的NPU，也就是图形处理能力，还是单纯为了算力或者CPU呢？

Were they buying it for its NPU based, you know, the graphics processing they have in there or were they buying it for just compute or a CPU, right?

所以，如果你回头想想人们为什么买Mac Mini，这个问题能告诉你一些东西。

So that question, if you revisit and ask yourself why people bought the Mac Mini, that will tell you something.

原因就是，人们不想在自己现有的笔记本电脑上运行OpenClaw，因为那等于让它访问你的整个电脑。

And the reason is people didn't want to run OpenClaw on their existing laptops because you're giving it access to your whole computer.

这很危险，因为你不知道它在访问什么。

And that's dangerous because you don't know what it's accessing.

你不知道它会拿你电脑上的信息做什么。

You don't know what it's going to do with the information on your computer.

它会不会访问敏感信息，比如医疗记录？

Does it access sensitive information, medical records?

你完全无法控制。

You have no control.

所以想法是，如果我买一台单独的机器，一台Mac Mini，专门用来干这个，在上面装OpenClaw，然后把它当作我的代理电脑。

So the idea is, what if I could buy a separate machine, a Mac Mini, keep it only for this purpose, install OpenClaw on it, and then have that as my agent computer.

然后在那台机器上，我不放那些敏感信息。

And then there I don't put all the sensitive information.

我只给它它想要的东西。

I give it only what it wants.

然后它就一直运行。

And then it works all the time.

现在，这种设置的问题，凡是运行过家庭服务器的人都知道。

Now, the problem with that setup, and this is familiar to anybody who's run a home server setup.

比如我就运行了大概十年。

Like I've run one for like ten years.

但保持它一直运行真的很麻烦，因为你要负责电力，对吧？

But keeping that up and running is a real big pain because you are responsible for power, right?

比如如果停电了，或者你的电脑，我不知道，电源的某个部分烧了或者出了什么事，电就断了。

Like if there is an outage or if your computers, I don't know, like some part of the PSU blows up or something happens, power goes down.

或者如果你的网络断了，那也是个问题，因为你没法连到家里的电脑。

Or if your internet is out, that's a problem because you don't have connectivity into your own machine at home.

但你还得设置网络，比如你像我这样在外面旅行的时候，怎么访问家里的电脑呢？

But you have to set up networking, like how are you going to access your PC at home when you're traveling like I am right now?

你得设置某种VPN网络。

How are you so you've got to set up some kind of a VPN network.

就像我们，你知道的，你和我经常用Tailscale来处理这些事。

Like we you know, you and I set up tail scale for all this stuff all the time.

这对每个人来说都不简单，比如如果我告诉我妈，嘿，你只需要买这个Mac Mini，然后设置Tailscale，再买个UPS防止断电，还要有备用网络以防万一。

It's not that simple for everybody to go off and be like, hey, if I tell my mom, hey, all you need to buy is this Mac Mini and then set up tail scale and, you know, set up a UPS so power doesn't go down and have some fail over Internet so nothing happens.

她肯定会说，算了，我搞不定。

She'll be like, I'm I'm out.

这些东西我用不来。

I can't use this stuff.

你能保留你的Agentic PC吗？

Can you keep your Agentic PC?

我就回去用我的普通笔记本电脑了。

I'll just go back to my regular laptop.

对，百分之百。

Yeah, a 100%.

是的。

Yes.

说真的，对99%的人来说，关键就是简单易用。

It is all about the easy button for 99% of people, seriously.

所以问题就在这儿。

So that was the problem.

所以人们买这些电脑，就是为了把环境隔离起来，给智能体什么的留一个独立的空间来运行。

So people are buying these computers just to sandbox their environment and have its own, you know, area for agents and stuff to run.

当时我自己也想了想，觉得，等等，做这件事的正确方式到底是什么？

At that time itself, I thought about it and thought, like, wait, what is the right way to do this?

比如，是不是非得买台电脑才能干这事儿？

Like, do you have to buy a PC and go do this?

不是。

No.

当时呢，你也可以租一台VPS，就是在服务器上租一个小虚拟机，然后在上面装OpenClaw。

At the time too, what you could do was you could get on a VPS, you could rent out a little virtual machine on a server and you could install OpenClaw on that one.

然后你就可以直接在云端跑，这其实是个更好的方案，因为你不用真的去买一台，大概六七百美元的机器，那台机器之所以那么火，就是因为大家都在买。

And then you just go off and run it on the cloud, which is a much better solution because you don't actually have to buy us, I don't know, $600.800 machine, which got so much more attention because like everybody's buying it.

所以就像一阵热潮，价格也涨上去了。

So it was like a craze and prices went up and all that.

除非你想在本地跑模型，因为确实有人买了Mac Mini，还是高配版，内存拉满的那种。

Assuming you don't want to run models locally because some people actually bought the Mac Mini, the high version with the maxed memory version.

现在回头看，这决定挺明智的。

Great decision in hindsight.

我觉得现在把内存拆出来卖，都能赚一笔，比整机还值钱，不过就当投资吧。但如果你想在本地跑模型，那也行，不过我觉得不是所有模型都能在本地跑，它们太大了。比如你想用前沿模型，像Fable那种级别的，配合你的智能体，那根本不可能。

I think you could pull out the memory and sell it for a profit compared to the cost of the machine itself now but consider it an investment but yeah if you wanted to run models locally you could do it right but I don't think models could all be run locally they're really big say you want to run a frontier model like the Fable class model with your agents, there's no way you could do that.

对。

Yes.

顺便说一下，我有一台智能体电脑，我确实在本地跑一些模型，但显然都是些小任务，不是前沿任务，包括音频处理，基本上就是转录，语音转文字。

I have an agent computer by the way, and I do run a few models locally, but it's obviously smaller things for the non frontier tasks, including audio, basically transcription, voice to text.

所以我会把这次播客用那台机器跑一遍，然后发到我们的Semi Doped通讯里，给大家提供文字稿。

So I'll run this podcast through on that when I post it to our Semi Doped newsletter to give the transcript to people.

这种简单的事情就可以在本地机器上跑。

That's like the simple kind of thing that can run on a local box.

但对大多数人来说，当他们想启动智能体并做事情时，我觉得有很多工作你会想要Opus或Fable那种质量的前沿模型。

But for most people, when they wanna spin up agents and they want to do things, I I think that there's a lot of work that you would want an Opus or a Fable quality, you know, frontier model.

对，这取决于工作负载，你可能想要更高质量的模型。

Yeah, it depends on the workload, you know, you might want a better quality model.

那么正确的做法是什么呢？

So what is the right way to do this?

现在，Grok Bot就是做所有这些事情的。

Right now, that's what Grok Bot does all of this.

它为你做所有事情，因为你不需要真的去弄，当你拿到Grok Bot时，这不是推销，他们也不是赞助商什么的，但我想说的只是它解决了什么问题，我不是建议任何人去买，也不是在推销这个产品。

It does all of this for you because you don't actually have to when you get Grok Bot This is not a sales pitch and they're not really a sponsor or anything But what problem it solves is all I'm trying to articulate I'm not suggesting anybody go and I'm not shilling this product.

好，那别买它。

Okay, so don't get it.

其实，它一个月要200美元。

Actually, it's $200 a month.

大多数人可能不会买。

Most people probably won't get it.

对。

Okay.

它非常贵。

It's very expensive.

所以，对。

So Right.

不是说你应该买。

Not saying you should.

但Grok Bot解决的问题是，你只需要拿到一个集成了所有功能的虚拟机。

But what this what Grok Bots solves is you just get the virtual machine with all the integration tied in.

安装的时候它会问你，嘿，你想要哪些集成？

It asks you when you install like, hey, what integration do you want?

我说，我想要我的谷歌日历。

I said, want my Google Calendar.

我想要我的谷歌云端硬盘。

I want my Google Drive.

你知道，这些东西你只要说，加，加，加就行了。

You know, some of these things you can just say, add, add, add.

然后所有这些都会自动装好，随时可以用。

And all those things are natively installed and ready to go for you.

你什么都不用做。

You don't have to do anything.

比如有Canva，有Figma，外面工具太多了。

Like there's Canva, there's Figma, there's so many tools out there.

你不需要一个个去设置。

You don't have to set up each one.

所以你在你的Grok Bot实例里设置好之后，它就可以在你所有的代理里使用了。

So you set it up there into your Grok Bot instance and it's available across all your agents.

然后它运行在虚拟机上，你什么都不用操心，包括那台虚拟机的安全，因为那是Grok的事。

And then it runs on a virtual machine and you don't have to be responsible for anything, including the security of that virtual machine because that's Grok's problem.

我希望他们有专家来处理这些。

I hope they have experts taking care of that.

我现在可不想变成网络安全专家。

I don't have to become like a cybersecurity expert now.

我是说，我也不太清楚。

I mean, I don't know.

我以为埃隆把X的人都裁了，或者大幅缩减了，不过算了，你继续。

I thought Elon fired all the people at X or slimmed But it way anyway, carry on.

当然，我相信他们肯定有……

Sure I'm sure they've got their

也许有办法，但他们没有。

Maybe ducks in they don't.

好的。

Okay.

行吧。

Fine.

那就部署Fable来当你的安全专家吧。

Then deploy Fable to be your security expert.

这就对了。

There you go.

这就对了。

There you go.

奥斯汀，你得学会创造性地解决问题。

Got to be a creative problem solver, Austin.

对，随便吧。

Yeah, whatever.

就像，这一切都不行。

Like, all this is no.

你什么都不用担心。

You don't have to worry about anything.

你会得到一个虚拟机，沙箱安全据说是有保障的。

You get a virtual machine that sandbox security is supposedly ensured.

然后你所有的工具都在Star里。

Then you've got all your tools in Star.

你运行你的代理，效果很棒。

You run your agents and it's amazing.

所以现在我们来看CPU的问题。

So now we go to the question of CPUs.

那我们还需要它们吗？

So do we need them?

让我们现在深入探讨一下，为什么我们真的需要CPU？

Like let's now go into the question of why do we really need CPUs?

因为在这整个过程中，我们甚至都没提到GPU。

Because we never even mentioned GPUs in all of this stuff.

我们确定做推理要用GPU，但自从智能体出现以来的这六个月里，我们其实根本没怎么聊过CPU到底为我们做了什么。

We are sure for doing the inference we want to do GPUs, but we really never haven't even touched upon what CPUs have really done for us in the last six months since agents showed up.

我觉得这正是我们现在应该深入探讨的话题。

And I thought that's what we should like dig into now.

那我们就来聊聊吧。因为从你举的例子来看，我们之前聊过在笔记本电脑上运行Cloud Code。

Let's dig into it because from your example, okay, so we talked about running Cloud Code on your laptop.

所以那里会有CPU在做一些工作，同时也会调用云端的GPU。

So there's going to be work done on the CPU there as well as calls to a GPU in the cloud.

我们还聊过在Mac Mini上运行OpenClaw。

We talked about running OpenClaw on a Mac Mini.

所以现在你桌上又多了一个CPU。

So now you've got another CPU on your desk.

不是你的笔记本电脑，而是你的Mac Mini。

It's not your laptop, it's your Mac Mini.

它有自己的独立运行环境。

It's in its own little contained environment.

它有一个CPU，但还是得连到云端去用GPU。

It's got a CPU and yet it's still running to the going to the cloud for the GPU.

然后，当然，你现在说的是用Grok Bot，它实际上是在云端启动一个虚拟机。

And then, of course, now you're talking about using Grok Bot, which is actually spinning up a VM in the cloud.

所以就是云端的CPU。

So CPU in the cloud.

所以就是云端的CPU和云端的GPU对话，然后把结果传回给你。

And so it's a CPU in the cloud talking to a GPU in the cloud and then communicating results back down to you.

所以实际上——我们待会儿会详细讲——我想说的是，有不同的工作在发生。

So there's actually and we'll get into this, but what I'm alluding to is there's different work happening.

在某些情况下，是在你本地的CPU上进行的。

In some scenarios, it's happening on your local CPU.

有些是在你家服务器的CPU上，还有些是在租来的云端CPU上。

Some it's on your home server CPU and others it's on a rented cloud CPU.

所以显然这里面有一些有趣的潜在影响。

And so there's obviously interesting implications.

好，那我们进入正题吧。

But, yeah, let's get into it.

先聊聊CPU到底在做什么，GPU又在做什么？

Let's first talk about like what what do the CPUs even do versus what is the GPU doing?

节目开始前，我们聊到了一个很好的类比。

Before the show, were talking about a nice analogy we could come up for this.

对，你来解释一下这个类比吧，因为这是一个很好的框架，能帮助大家理解GPU和CPU各自的位置，以及我们以后该怎么看待这个问题。

Yeah, I'll let you explain that analogy because this is a very nice framework for which people can think about where GPUs lie, where CPUs lie and how we should think about this going forward.

好的，没问题。

Yes, okay.

我觉得最能体现价值的地方就是GPU。

So ultimately, I think of a good way of like where the value happens is the GPU.

GPU就像个天才，你懂的。

The GPU is like the genius, you know.

它就像个什么都懂的博士，当然，你并不总是需要这样的天才。

Now it's like the p h the the genius that has a PhD in everything, which of course you don't always need.

但是GPU就像那个天才，不过说到底，一旦那个大脑，那个天才，想好了所有需要做的工作——我的意思是，这就是我在Fable做的事情，我在说，给我你深思熟虑的方法或者opus two。

But the the GPU is like the genius and then but at the end of the day, once like the brain, the genius, like thinks of all the work that needs to be done, which I mean, this is what I'm doing with Fable is I'm I'm trying to say like, you know, give me your very thoughtful approach or opus two.

然后一旦决定了需要做什么工作，最终，这些工作，取决于工作量，但很多工作其实可以交给助手，那些小执行者，小任务执行者。

And then once it is decided, like what work needs to be done, ultimately, of this work, it depends on the workload, but a lot of the work can actually be given to, like, assistants, like little doers, little task carry outers.

所以最终，你知道，我觉得GPU就像那个天才，负责所有思考和分配工作，然后CPU就像助手，一旦被告知要做什么，他们就可以去执行天才告诉他们做的所有工作。

So ultimately, you know, I think about the GPU is like the the genius that is doing all the thinking and delegating work and then ultimately the CPU as the assistance that once told what to do, they can go carry out all the work that the genius told them to do.

所以理想情况下，你不想让天才做所有的工作。

So ideally, you don't want a genius doing all the work.

比如，天才在思考。

Like, the genius is thinking about it.

天才每小时花很多钱来想出这些绝妙的主意。

Genius has paid a lot of money per hour to come up with these brilliant ideas.

你不想让天才去，我不知道，把文件放进抽屉里，或者你不想让天才去从抽屉里的文件中查找信息，甚至不想让天才去取邮件。

And you don't want the genius going off and, I don't know, putting paperwork in the drawers or you don't want the genius looking up some information from a file that's in the drawers or don't even want the genius going out to fetch the mail.

理想情况下，在理想的世界里，你希望让那个绝对的天才，也就是GPU，只做它最擅长的事情。

Ideally in an ideal world, you want to have the absolute genius, AKA the GPU, just doing what it does best.

其实，这个比喻还可以再细一点。

Actually, there's one We can break this analogy down into a little bit more detail.

比如说，如果所有工人都是CPU，像CPU核心那样各干各的，我觉得我们也可以把这种工人细分一下。

So if all the workers are like CPUs, like CPU cores doing their thing, for example, I think we can break down what kind of worker that is too.

因为有一种东西叫头节点CPU。

Because there is something called the head node CPU.

头节点CPU的比喻就是天才的助理。

And the analogy to the head node CPU is an assistant to the genius.

这个助理是给天才——也就是GPU——服务的，它的工作就是不停地给天才派活。

The assistant to the genius, which is the GPU, the job of this assistant is to keep handing the genius work.

如果天才需要咖啡，你就给他咖啡。

If the genius needs coffee, you give the genius coffee.

如果天才需要水，你就给他水。

If the genius needs water, you'll give it.

天才不能停下来，因为天才一停，你就在亏钱。

The genius doesn't stop working because you're losing money if the genius stops working.

你要确保天才完成这项任务后，下一批材料已经准备好，让天才继续工作，主机节点CPU就会说，给您，先生。

You want to make sure that once the genius is done with this task, the next folder of materials is ready for the genius work to be done and the host node CPU would be like, Here you go, sir.

这是你接下来要做的事。

This is what you've to do next.

天才就会说，走起。

A genius is like, Let's go.

干吧。

Let's do this.

所以这就是主机节点，我觉得它的单核性能必须非常快，因为它要不断监控GPU在做什么。

So that's the host node and I'd say it has to have a pretty fast per core performance because it's constantly monitoring what the GPU is doing.

你完成了吗？

Are you done?

你完成了吗？

Are you done?

还没有。

No.

好的。

Okay.

比如说，咱们把这个准备好。

Like, let's get this ready.

让它随时待命。

Let's keep it ready.

GPU一完成，这边就得马上交接。

As soon as the GPU is done, this has to hand it in.

就是说，不能有延误。

Like, you know, no delays.

不能浪费时间。

No wasting time.

别浪费天才的时间。

Do not waste the genius time.

所以这就是主节点CPU的职责。

So that's the host node CPU.

是的。

Yes.

我很喜欢。

I love it.

我很喜欢。

I love it.

这说明你既想要高主频或者响应快的CPU，同时也希望CPU和GPU之间，也就是助手和天才之间的通信尽可能快。

It's it shows that you want like a high clock speed or a very responsive CPU, but also obviously you want like the communication between the CPU and the GPU, between the assistant and the genius to be as fast as possible.

所以，比如说，理想情况下，助手不应该在另一栋楼里，然后他们得过来签到，再走回自己的楼，然后再回来签到。

So that there's no For example, ideally, wouldn't have the assistant be in a different building, and then they have to come over and check-in and then walk back to their building and then come back and check-in.

你希望他们就在同一个房间里，直接互相交流。

You want them, like, in the same room talking with each other.

而且，再深入一层，这些主机节点在一致性主机和标准主机之间是有区别的。

And and so maybe to take it, like, a layer deeper, there's differences between a coherent host and a standard host with these host nodes.

比如一致性主机就像Grace Blackwell，GPU和CPU可以通过共享的高速网络通信，比如NVLink CTC；而标准主机可能只是用一些Hopper，然后放一个x86处理器在主板上，但还是通过PCIe通信。

So like a coherent host would be like Grace Blackwell, where you've got the GPU and CPU and they can talk over a shared really fast network, so like NVLink CTC, versus like a standard host where maybe you have just some hoppers and you take an x 86 processor and you put it on the board, but it's still talking over PCIe.

那就有点像他们分别在独立的楼里或房间里。

So that would be a little bit more of an equivalent to like, they're in separate buildings or separate rooms.

所以芯片到芯片，也就是GPU和CPU之间用的C2C协议，是一种一致性协议，你知道，一致性CPU就像一个人就站在旁边，

So the chip to chip, the C2C protocol that the GPU and the CPU uses, coherent protocol that, you know, the coherent CPU would be somebody who's standing right next to

然后

the and

看着你，就像说，哦，给你。

looking over, I'm like, oh, here you go.

给你。

Here you go.

你懂吧？

You know?

就在那儿。

There.

对，他们如果，是的，就是这样。

And if they yeah, exactly.

如果他们有内存一致性，就像，我能看到，我是助手。

If they have like memory coherency that they it's like, I can see the I'm the assistant.

我能看到天才的笔记，我也可以在他的笔记上写点什么。

I can see the geniuses notes and I could like jot down on his notes too or something.

所以我在想HBM在这个比喻里是什么，这里我们比喻得有点过头了，你不应该这样，但我们还是继续吧。

So I was thinking of what HBM is in this analogy and this is where we push it too far and you shouldn't but we'll do it anyway.

但我把HBM想象成天才书桌旁的一叠纸。

But I think of HBM as a stack of papers next to the geniuses desk.

天才不会从椅子上站起来去柜子里拿。

The genius is not going to get up off the chair and go and get it from the cupboard or anything.

HBM就是桌上的那叠纸，天才可以直接拿过来继续工作，用完再放回去。

HBM will be the stack of papers on the desk and the genius can just take it and keep working on it and putting it back on or whatever.

内存一致性来自于助手也能接触到桌上同一叠纸，因为他们就在同一个房间里。

Coherent memory coherency comes from the assistant having access to the same pile of papers on the desk because they're in the same room.

所以如果助手想的话，也可以直接访问内存。

So if the assistant wants, can also access the memory directly.

对，对。

Exactly, exactly.

当你把视角拉远来看，这个CPU助手，也就是主机节点，它的任务只有一个，就是让GPU吃饱。

When you zoom out to your point, this CPU assistant, this host node, it has one job and it's keep the GPU fed.

这最终就是主机CPU要做的事，保证GPU有活干。

And that's ultimately what the host CPU needs to do, keep the GPU fed.

所以这就引出了一个有趣的问题。

So then it starts to raise the interesting question.

这在ChatGPT时代很有道理，那时候一切都只是个聊天机器人。

That made a lot of sense in the chat GPT area where everything's just a chatbot.

就像，嘿，我在问问题。

And it's just like, yo, I'm asking questions.

然后就是，只要让天才吃饱，他就能回答问题。

And it's just like, just keep the genius fed so he can respond to questions.

现在突然有了这么多额外的工作，比如天才在吐代码，你需要编译这些代码，看看能不能编译通过。

Now all of a sudden, when there's all this extra work to be done, like the genius is spewing out code and you need to compile that code, see if it compiles.

也许你需要运行一下代码。

Maybe you need to run the code.

也许这个天才需要从50个不同的来源获取信息。

Maybe the genius is like, I need information from 50 different sources.

去查SEC文件，去谷歌搜索，或者在网上搜，做所有这些事情。

Go hit SEC filings and go Google the web or search the web and go do all this stuff.

现在突然之间，问题变成了：那个CPU助手能去获取所有这些文件吗？

Now all of a sudden, the question is, can that CPU assistant go fetch all these filings?

它能处理所有这些数据吗？

Can it take all this data?

它有足够的内存容量和带宽来分析所有这些数据，并保持GPU的供给吗？

Does it have the memory capacity and bandwidth to analyze all this data and keep the GPU fed?

还是说，对于站在它旁边的那个助手来说，这工作量太大了？

Or is this too much work for that assistant standing right next to him?

我们是否需要一大群其他CPU来帮忙？

And do we need an army of other CPUs to help?

所以这就引出了一个问题，那是什么？

So this brings us to the concept of what is it?

我不知道。

I don't know.

有没有一种叫智能CPU的东西？

Is there a thing called an Agentic CPU?

我不知道。

I don't know.

那可能是个营销术语，但我觉得在这个语境下说得通。

That's like a marketing term probably, but I think it makes sense in this context.

那么，我来谈谈我对智能CPU的看法。

So, you know, I'll give you my take on the agentic CPU.

其实，我最近刚和AMD录了一期播客，他们的想法和我们非常一致。如果你把视角拉远，抛开这个比喻，回到过去云时代的工作方式，显然，我们有很多通用CPU。

Actually, I just recorded a podcast recently with AMD and they were aligned and thinking very similarly, is if you zoom way out, sort of breaking this analogy, just going back to how things used to work, back in the cloud era, obviously, we had all these general purpose CPUs.

它们支撑着我们在云端使用的所有SaaS产品，对吧？

And they back all of the SaaS products that we use in the cloud, right?

所以现在，当我们搭建起所有这些GPU，就为了问它们那些小聊天问题的时候，它们当然不一定会去跟一个已经在运行API服务器或者我们公司数据库的CPU通信。

And so now as we're standing up all these GPUs so that we can ask them our little chat questions, of course, they're not necessarily gonna go communicate with a CPU that's already running an API server or running my database for my company or something.

就像，那个CPU已经有自己的活儿了。

Like, that already has a job.

而且那个CPU已经针对特定的工作负载做了优化，比如快速运行大型数据库，或者快速运行大量小型API服务器之类的。

And that CPU is kind of already shaped to fit the workloads of like, I run big databases quickly or I run tons of small little API servers quickly or whatever.

所以问题就来了，好，现在快进一下。

So then the question is, okay, now fast forward.

如果我们有所有这些通用GPU，它们已经有自己的工作了，已经在运行SaaS产品、Web服务器或者数据库之类的。

If we've got all these general purpose GPUs that already have jobs, they're already running SaaS products or web servers or databases or whatever.

然后现在我们有了这些新来的“天才”，它们旁边有个小助手CPU，负责给它们喂数据。

And then now we've got these new geniuses that are standing up and they have a little assistant CPU next to them trying to keep them fed.

但如果现在有工作要溢出来，因为主机节点的CPU忙不过来，那么问题就是，这些工作应该落到哪些CPU上，它们应该放在哪里？

But if now there's work that's going to spill out because it's too much for the host node CPU, then the question is, what CPUs should those fall on and where should they live?

所以最近出现了一个术语，叫“Gen 6 CPU”，这更多是说，嘿，这些CPU是专门用来做这种代理工作的，就是那些主机CPU本来能做但太忙了没空做的溢出工作，因为它们要忙着喂饱助手或者喂饱那个“天才”。

And so there is this term of art that has come up lately, a Gen six CPU, which is more about, hey, these are CPUs that are dedicated to doing this agentic work, all this spillover work that the host CPU could do, but it's too busy keeping the assistant fed or it's keeping the genius fed.

所以应该有专门的机架来放这些CPU，专门干这个活儿。

So there should be racks of CPUs that are dedicated to this task.

但当然，这又引出一个问题：这些CPU应该是什么样子的？

But of course, it then raises the question, well, what should those CPUs look like?

它们应该跟主机一样吗？

Should they look like the host?

还是应该跟助手一样？

Should they look like the assistant?

或者更偏向通用型？

Or should they look more like the general purpose?

还是说它们需要有自己的形态？

Or do they need to have their own shape?

所以这就是这种新型的、介于两者之间的智能体CPU的大致框架。

So that's kind of like the framing for this new kind of middle ground agentic CPU.

智能体CPU。

Agentic CPU.

对，所以就像你说的，你不想用主机CPU来做天才要求做的这些工作。

Yeah, so you don't wanna use the host CPU, like you said, to do any of this work that the genius is asking.

比如，嘿，去网上搜一下。

Like, hey, go search the internet.

不行，因为一旦主机节点去做这种事，它就不再给GPU和天才供数据了，然后整个系统就崩溃了。

No, because once the host node goes off and does something like this, it stops feeding the GPU and the Genius and then the whole thing goes down.

那还有什么意义呢？

Like what's the point?

现在天才闲置了，这是最糟糕的情况。

Now the Genius is idle, which is the worst case scenario.

所以你需要一种不同的CPU来做这些事情。

So you need a different kind of CPU that does this stuff.

甚至不一定非得是不同类型的CPU。

It doesn't even have to be a different kind of CPU.

功能上它是不同的CPU，但你可以用和主机节点一样的CPU。

It's in function, it's a different CPU, You could use the same CPU like the host node CPU.

它可能不需要我们之前提到的那种一致性，因为它不是一直在和天才交流。

It may not need the coherency that we spoke about because it's not talking to the genius all the time.

对。

Yes.

但除此之外，它可以是同一个CPU。

But otherwise, it could be the same CPU.

但理想情况下，你可以这样想：比如我们拿一个CPU，比如说有128个核心之类的。

But ideally what you can think of this is like let's take a CPU with like I don't know let's say 128 cores or something.

所以你可以把所有这些核心想象成办公楼里的平面图，然后你可以画一些小方框，说，好的，这四个核心是财务部门。

So what you can do is you can think of all the cores and as like a floor plan of an office building and then you can kind of draw little boxes and say, okay, these four cores are the finance department.

这四个核心是我的研究小组。

These four cores are my research group.

这四个核心是我的设施团队，随便什么，对吧？

These four cores are like my facilities team, whatever it is, right?

然后你就像有了一个小公司，天才在说，嘿，我需要这些事情完成。

And so what you do is then somehow you've got yourself a little company and the genius is saying, Hey, I need this stuff to be done.

主CPU会在房间里喊：嘿，团队，我需要你们去把昨晚的SEC文件抓取过来。

And the host CPU is gonna yell across the room and say, Hey, team, I need you to go and scrape up the SEC filings from last night.

去办吧。

Go and do that.

那些人会说：好的，没问题。

And those guys will be like, Okay, cool.

我们会去做的。

We'll do that.

然后那四个核心组成的团队就会去干活，说：好的，我来处理这些。

And the team of four cores will go off and be like, Okay, I'm going to do this stuff.

他们会向主节点CPU，也就是助手汇报，然后助手再把结果交给天才，也就是GPU。

They'll report back to the host node CPU, the assistant, who will then feed it to the genius, the GPU.

这种场景，就是多核CPU，比如AMD的256核或者Intel最新的288核CPU。

This scenario, a multi core CPU that is either like one of the AMD's two fifty six core or Intel's more recent two eighty eight core CPU.

这些CPU很棒，因为你可以分配很多小部门去做不同的功能，或者你也可以有更大的团队。

These are great because you can assign a lot of little departments across doing different functions or you could have basically bigger teams of people.