저자: Nick·Kasey · 외국인
요약 보기
브로드컴의 XPU(ASIC)가 엔비디아 GPU를 대체한다는 미디어 논조를 바로잡는 영상. 브로드컴 AI 반도체 매출은 53억 달러로 반도체 매출의 절반 이상이며 전년 대비 63% 증가, 4분기에는 66% 증가 전망. 잉여현금흐름은 매출의 약 43%. XPU는 특정 워크로드용 맞춤 칩으로 GPU와 완전한 대체 관계가 아니며, 하이퍼스케일러는 GPU와 XPU를 병행하고 실제 경쟁 상대는 CPU 기반 범용 서버이기도 하다. 여러 기업 바스켓 보유가 낫다고 제시.
다듬은 글
Hey everyone. Welcome back to Chip Stock Investor. It's Nick and Kasey. And after looking at, a ton of media headlines and, commentary about how Broadcom's XPUs are suddenly going to rain on NVIDIA's GPU Party. We thought we should maybe address this topic, probably not one last time, but let's address it again because after two, three years of covering this, it seems like there's still a lot of confusion about the difference between an XPU, a type of application specific integrated circuit, and NVIDIA's slightly more general purpose GPUs.
We're also doing a deep dive into Al Chip, which could be a massive disruptor for Marvell Technology. Just our semi insiders will get that research deep dive. You can check that out over on our website, chipstockinvestor.com. Now to introduce this topic, we should probably just start with Broadcom's revenue segments this chart provided by our friends [email protected], you can do great financial visuals including key performance indicators for companies like Broadcom to get a better understanding of where they're actually making their money.
Here in this chart, we have Broadcom's Semiconductor Solutions highlighted in red and infrastructure software, including, of course, VMware highlighted in blue. You can make cool charts like this one by getting a fiscal.ai membership. 15% off with the link and, the description below or here on the screen, even for that enterprise plan, for some of you pros out there, 15% off fiscal.ai/csi.
Broadcom's AI specific semiconductor revenue was 5.3 billion, over half of their semiconductor revenue, in the most recent quarter, and that was up 63% year over year. It's expected to be up even more 66% year over year in fiscal Q4, which ends in October. One more financial slide here. As Broadcom grows that AI semiconductor business and of course, as it further integrates VMware into its enterprise software platform, take a look at those free cash flows.
This time, free cash flow is in blue. GAAP net income in red, free cash flow, about 43% of revenue in this latest quarter, that ended on August 2nd. Pretty incredible stuff for Broadcom as they continue to claw their way back up towards that 50% free cash flow margin they reported prior to the VMware acquisition, almost two years ago now.
But Kasey, let's get back to the original question. What is the difference between an asic, like an XPU, that's part of Broadcom's AI semiconductor business, just part of it. They also have a lot of, networking, like the newly announced Jericho four. But let's talk about this because that seems to be the big buzz right now is suddenly, uh, I don't know why this is surprise, but Broadcom is a big deal.
Their XPUs are competing with Nvidia. So let's maybe set the record straight. Is that true? And I think to answer that question, we can talk about what actually an XPU is versus what a GPU is because there is a difference. We have this broken down in three parts. Merchant on the left, GPU or graphic processing unit in the middle and ASIC or application specific integrated circuit on the right.
Merchant chips those are for general purpose work. Quote unquote, off the shelf. And just until a couple of years ago, this is actually what CEO Hock Tan thought would be the winner for Broadcom. And up until a couple years ago, it really was their biggest segment off the shelf silicon. Yeah, it's ready to go plug and play.
The customer doesn't need to do anything with it, but now enter the GPU especially three years ago from Nvidia. And these are type of merchant silicon, but they can also be used for specific applications so they can be customized. And that's the beauty of NVIDIA's Cuda developer ecosystem, right? All those software algorithms that can help customers speed up the development of whatever workload they're working with using gpUs accelerated computing.
So let's call that somewhere in the middle. And then on the far right, you have the application specific chip, including the XPU or accelerated processing unit from someone like Broadcom. So this more customized product, the ASIC or XPU from a company like Broadcom, certainly sounds like it might be a threat to Nvidia.
And to a certain extent, perhaps that is true. But it its own oversimplification just to simply say, oh, Broadcom is now, a big Nvidia competitor as the big hyperscalers, the big cloud operators try to cut Nvidia out of the mix. That's not exactly what's happening here. We're gonna oversimplify this by talking about tools that you have in your kitchen.
You have general purpose tools, and then you have something that might be really, really niche application specific. For example, Kasey came up with the oven versus the toaster. Hopefully this works everyone. it works. The oven, you can cook or bake all sorts of things in. You can roast a Turkey if you want, or you can bake a cake, but the toaster is definitely more application specific.
It's for one purpose. The GPU can do all sorts of things, including what the ASIC could do, including XPU. But the XPU is designed for one very specific, one niche thing. And the reason why this has become a thing, it's because, the semiconductor industry has disaggregated over the years, and there are lots of companies that specialize in different parts of the supply chain, including Broadcom, a big fabless chip designer, and it owns lots of IP and patents.
But the other reason this has begun to work is we have these large hyperscalers data center operators like Google and Google Cloud, that have such massive workloads operating in their data center that may be one very specific workload, one very specific software operation could benefit by custom designing something like the XPU, or as Google has called it since 2015, the TPU, the Tensor Processing Unit.
We're now on the sixth Gen TPU or Trillium. Kasey, walk us through, generally speaking how this would work between Broadcom and Google slash Google Cloud. This is where Broadcom's IP comes into play. Their ASIC IP or intellectual property. They may provide custom design for a specific software for Google Cloud.
Google and Broadcom then take that design and send it to a manufacturer like Taiwan Semiconductor Manufacturing, who then manufactures that XPU and the XPU is then installed in maybe the data center where it provides that very custom workload for Google. And so to simply answer this question is Broadcom big new competition for Nvidia?
Uh, well, kind of, kind of. Generally speaking, these big hyperscalers aren't purchasing Nvidia GPUs for those ultra specific cloud or, or software needs. Their general purpose. they can do things like train , new AI frontier models. They can operate, software. They're doing lots of research and development with those NVIDIA systems.
And the ASIC is probably a very large tried and true, all the bugs are worked out software application. That application has gotten so big that, okay, now it makes sense to just go get an asic. And oftentimes this isn't competing directly with Nvidia, who it's actually competing with is sometimes just general purpose merchant silicon, for example, like a CPU based server from Intel or AMD.
So a lot of this narrative about Broadcom competing with Nvidia. Yeah, it's kind of true, but it's actually just as much true, if not even more so that the ASIC for these big cloud and data center operators is competing with the merchant silicon, with the very large installed base that exists out there in data centers that is CPU based computing.
So in this situation it's a this and that, not this or that. Could be. We'll show you this slide that has a sampling of competitors in this field who maybe provides networking chips or custom asics and merchant silicon as well. And I think this illustrates why it might be better to own a selection or a basket of these companies versus betting the farm on maybe Broadcom or Nvidia or even Marvell and thinking that this is a winner takes all situation.
Yeah, it's a massive market at this point. Hundreds of billions being spent on this infrastructure. There's a lot of room for a lot of companies to win, and it's not just a, as you said this or that situation for these hyperscalers. They have very large, very complex data center infrastructure installed, and it's getting more complex as every year goes by.
Okay, let's close this out. We hope to see you over on Semi Insider. We already did our Credo update last week, and as you mentioned, we're gonna be talking about Al Chip over there, this week as well as some other hot news Synopsys, earnings first report, synopsys plus Ansys. We'll also have a discussion on Oracle and some more cybersecurity on our live events.
Check it out. Semi insider on chipstockinvestor.com. And just a quick reminder, hit the subscribe button if you haven't already. We really appreciate it and thanks for watching our non AI generated content. Thanks everybody. See you soon at Chip Stock Investor.
Comments 0