KETTLE This week, we learned a lot more about what’s going on inside the datacenter. GPUs may still be king but their exclusive reign on the rack is over. The taskbar's supremacy at the bottom of the desktop may be coming to an end too.
You can listen to the latest episode of The Kettle right here on this page, as well as on Spotify, Apple Music, or YouTube. Those platforms also let you subscribe to Kettle, so you are always notified when the latest episode goes live.
Register Systems Editor Tobias Mann joins us to talk about why so many vendors are turning to alternate silicon from the likes of Cerebras and Marvell to help power their inference. He also explains why Waymo is designing its own custom chips in robotaxis rather than continuing to use off-the-shelf silicon from Intel.
As new AI silicon moves forward, Microsoft goes back to the future by finally, after only five years of waiting, allowing users to move the taskbar in Windows 11. It also has belatedly started giving users more control over context menus and given them a closer look at what their neural processing units (NPUs) are doing in task manager. Register Microsoft Ecosystem Reporter Richard Speed weighs in.
A lightly edited transcript is below.
Avram (00:04) Hey everyone, welcome to this week's episode of The Kettle. I'm Register US Editor Avram Piltch, and I am joined today by Tobias Mann, our Register Systems Editor, and Richard Speed, our Microsoft Ecosystem Reporter. We're going to talk about some really exciting new developments in AI hardware and changes to Windows 11 that have been long overdue. So let's get started.
This is thankfully a quieter week than usual, maybe because we're nearing the end of August and we don't have to talk about the AI bubble popping or any kinds of economic disaster. We can actually talk about some new technological developments and geek out over those, which is a lot more fun. One thing that we've noticed is that a lot of AI companies seem to be moving away from traditional GPUs and are looking at different types of hardware. Tobias, I know you've covered this in a bunch of stories. Specifically, I wanted to get started by talking about Cerebras. Can you tell everybody a little bit about Cerebras and what they do?
Tobias Mann (01:24) Sure.
Cerebras has been around for quite a while. They've been poking at the AI infrastructure problem for a lot longer than large language models have been the dominant force in all of this. The company, founded and led by Andrew Feldman, differentiated itself pretty early on by, rather than building GPU-sized chips, their chips are the size of dinner plates or charcuterie boards, and this has fgiven them an early advantage during the ChatGPT days up until recently. Because they're not reliant on high-bandwidth memory, which is supply-constrained, their chips are so large that they can pack 44 gigabytes of super-fast petabytes-per-second SRAM on them.
[Cerebras founder Andrew] Feldman is notoriously not a fan of Jensen Huang or Nvidia. He previously labeled Nvidia as an AI arms dealer for continuing to push GPU sales into China
Tobias Mann
Starting around 2024, they launched an inference server that allowed them to serve tokens way faster than any GPU on the market. It doesn't matter how many GPUs you stack together, they could serve the model faster. However, they did have some challenges associated with the amount of compute that they had. A big shtick for them was that they could get more compute out of their chips using something called structured sparsity. It takes advantage of the number of zeros in the training sets. While that's great for training models, it doesn't do anything for inference. You would think that would have been a dead end for them, but they have found a path forward by partnering with other companies to do the compute-heavy side of inference for them. Earlier this year, they partnered with AWS for this purpose. Just last month, they partnered with AMD. Feldman is notoriously not a fan of Jensen Huang or Nvidia. He previously labeled Nvidia as an AI arms dealer for continuing to push GPU sales into China.
There has been this shift away from trying to do everything on one kind of chip, a GPU or custom AI accelerator, to distributing it across two. If that weren't enough, we were getting towards the end of the lifecycle for Cerebras' last chip, and just this week we got a glimpse of what its next generation is going to look like.
Avram (04:25) So, what is the difference with the next generation of Cerebras?
Tobias Mann (04:30) Well, the biggest differences come from the rack architecture. These chips are huge and super power-hungry. The last generation took fifteen kilowatts just to power the chip. To put that in perspective, modern Nvidia GPUs pull about 1.2 kilowatts, so they're quite a bit hotter. They used to be placed in these big monolithic chassis – think of a server that's half a rack tall, liquid-cooled, and had everything you needed to deploy it in that one box. But as it turns out, data center operators don't really like operating like that. They would prefer to be able to interchange power, cabling, and compute modularly. What we've gotten this generation with the CS-4 systems is a rack-scale system that includes three chips in a modular fashion. If a compute unit goes bad, the customer only has to swap the compute, they are not also replacing all of the power delivery equipment as well. Another big improvement with this generation is that the power delivery is so much better that Cerebras has doubled the clock speed on its chips. It's calling them a turbo variant, but in reality what they're doing is pushing twice as much power through them. So they've been able to double the performance, double the bandwidth of the memory, and double the I/O speed. They're effectively getting twice the compute out of the same silicon, and they've put three times as much in the rack.
Avram (06:24) Who's going to use Cerebras?
Tobias Mann (06:26) This is an interesting question because it has changed a little bit since Cerebras launched their inference service in 2024. At the time, Cerebras was mostly going after customers that needed to generate what are called premium tokens in the industry. It's a really silly way of saying, "Hey, I can generate hundreds or thousands of tokens a second," where a GPU system with maybe six dozen GPUs might only be able to do a few hundred per user. This is beneficial for things like coding assistants. One of Cerebras' largest customers to date was OpenAI. When OpenAI launched their Codex code assistant, they also launched a Spark model that ran specifically on Cerebras hardware because of how quickly they could generate tokens. Fast tokens are great for code assistants, but fast tokens are also cheap tokens. The faster you can generate tokens, the more work you can get done in a second, driving down the operating cost of the hardware and how much you have to charge to draw a profit.
At the end of last year, Nvidia acquired Groq for $20 billion. The entire purpose was that Groq did something very similar to Cerebras, achieving hundreds of tokens a second for things like coding assistants. By combining them with Nvidia GPUs, Nvidia was going to be able to do that for thousands of users at those speeds. It's high volume, high speed.
The rest of the industry didn't have an answer to this. So Cerebras has positioned themselves as the alternative to Nvidia/Groq for everybody else in the industry. AWS is one of the major partners for doing this, so expect to see services built atop AWS's AI services to rely heavily on Cerebras. I would also expect to see AMD, as they're starting to sell their own rackscale system that we looked at back at Advancing AI last month, I also anticipate we're going to see AMD pushing it as a kind of two-part deployment option to its customers, in order to drive higher speeds for premium tokens but also drive down operating cost of running inference services like Base10, Fireworks, or any of these other AI-as-a-service providers.
Windows is a supertanker – it takes a long time to turn that supertanker around
Richard Speed
Avram (09:23) Makes a lot of sense. We also see that Google is moving away from traditional GPUs, or at least getting another partner to help them with their TPUs, correct?
Tobias Mann (09:38) Yeah, that also happened this week. A little bit of context here: Google is a cloud operator. I don't think that's a surprise to anybody at this point. Every sufficiently large hyperscaler eventually ends up in this business. Google's customer base has long consumed traditional GPUs, and since around 2015, Google started developing and introducing its own custom AI accelerators called Tensor Processing Units (TPUs), primarily for internal workloads. This is something we've seen with a lot of cloud providers: as they scale up, many have advertising businesses or recommendation systems for search traffic where it's beneficial to have something that customers don't necessarily need as much of.
The open secret with all of these custom cloud accelerators, CPUs, etc., is that none of them are really built from scratch. Google has lots of talented chip engineers, but there's a lot of stuff that goes into an AI accelerator or CPU that is just boring, and if you put a bunch of engineers on this stuff, you're just reinventing the wheel.
These cloud providers have relied heavily on big IP houses like Broadcom to design the boring bits of the chip so they can spend most of their time designing the fun stuff that's going to make them money, whether serving advertisements or powering AI services. as was the case of Google. Broadcom was that partner for Google, that was publicly disclosed as part of an Anthropic deal.Broadcom started getting rather loud about being the IP provider for a lot of these chips. They also provided IP for Meta, that was publicly disclosed this year.
The sea change this week was that Marvell is now getting involved in all of this. Marvell is another IP house that has largely built its holdings through acquisitions of companies like Cavium, I believe Innovium is also in there if my memory is serving me correctly. They bought Celestial AI for photonics connectivity late last year in a multi-billion dollar deal. Marvell has long been a competitor to Broadcom in the networking space, but Broadcom is a massive monster of a company.
There's a reason we call them Broadzilla; their IP holdings are massive. Marvell has reached a point where it's now competitive and has IP for custom XPUs. Google could go to Marvell and ask them to design a TPU alongside Broadcom and then see which performs better. This is I think where things are going.
As anybody who's followed Broadcom long enough knows, once they have an effective monopoly, prices and licensing fees tend to go up, much like we saw with VMware not too long ago.
Avram (13:19) So Google is getting in on the action to do more to get custom silicon, and now we find out that robotaxis are getting special silicon as well, which I don't fully understand why they need. Perhaps you can explain, because Waymo is now designing custom silicon.
Tobias Mann (13:45) Yeah, this was one I wasn't expecting this week. Waymo disclosed that they had developed a custom machine learning accelerator specifically for autonomous vehicles. This is a market where there is no shortage of off-the-shelf options from companies like Intel Mobileye, Qualcomm, Nvidia, AMD. Basically, if you want to build a robotaxi or autonomous vehicle, you don't have to look very hard to find something off the shelf. We've seen companies like Tesla go down this road, but Tesla is a massive electric vehicle company that has made a lot of inroads through technological design and innovation around full self-driving, so it's less surprising there. Waymo is a lot smaller, but arguably a lot more successful in getting robotaxis on the road that actually work and are legal and have actual potential of getting you where you want, although getting stuck in a roundabout is always a hilarious sight to see.
For Waymo, the problem they're trying to solve is that these cars have a tremendous amount of sensor data streaming through them. There's very little time between seeing an obstacle on the road and taking evasive action to avoid it. While Waymo can have a human pop in and take control of the vehicle, that's not an ideal situation if you're hurling down the road at 30 miles an hour and someone steps into the street. By the time a human takes control, it's too late. Latency is a huge factor in how quickly it can spot an obstacle and hit the brakes or steer away. So Waymo has decided to build this chip. Avram, you mentioned why they would need this given all the off-the-shelf stuff that had been working for them. From what I can tell, they have largely been building these autonomous systems using FPGAs. FPGAs are fantastic for super-low-latency workloads. They're entirely programmable, but they're hard to program for, it's a very specialized group of people who can program for them, and they're not very compute-dense. To add more sensors, you have to add more FPGAs to the compute stack of these vehlcles. That consumes more power and compounds your development maintenance stacks.
I suspect Waymo's motivation for moving toward an ASIC in this case is to get better compute density that's easier to work with from a development standpoint, without missing out on the low-latency characteristics FPGAs are infamous for. There's a reason why all high-frequency trading is done with FPGAs today – if you want to take action very quickly based on a stimuli, they're very good for that. But ASICs can be just as good if they're designed for that purpose.
Avram (17:48) Makes a lot of sense. So I think what we're seeing here is that there's more diversity of hardware going into inference, going into the AI industry, so that we can get better power management, better performance, maybe better economics, easier to program. So a lot of a lot of interesting things happening here that are moving outside of the domain of just the big big hardware companies like Nvidia.
Speaking of big companies, I wanted to shift gears and talk about Windows 11, which is almost five years old now and still can't do some things Windows 10 or 7 could, like let you move the taskbar. I wanted to ask you, Speedy, as I read that we might finally be able to move the taskbar again. Is that true?
Richard Speed (18:57) That is true. Full disclosure: it's never bothered me personally having the taskbar on the bottom, but for some users they really care about this. And it's one of the things that Microsoft removed, it was in Windows 10, as you say, it was in Windows 7. And yet inexplicably when Windows 11 arrived, just well it's almost five years now, they'd taken it away. And the the reason for that was they basically rewrote the the interface they decided to drop moving the taskbar, and users have complained about it ever since. What's interesting is that it's now coming back – it has arrived in the Release Preview channel of the Windows Insider program, meaning it should be out publicly in the next few weeks or months...
What I think it speaks of is paying attention to users once again because this has been a thing people have been asking for. They didn't ask for Copilot everywhere, they didn't ask for ads everywhere. They just wanted to move their taskbar around like they always used to. And so finally that's happening. I mean it kind of echoes comments that came from the the Windows boss Pavan Davuluri ... early this year he said Microsoft would be rethinking how they put Copilot into things, they'd be rethinking things like the taskbar and that the fact that users want to do it to work a certain way. They'd be rethinking the Start menu. They'd be paying more attention. And I think when your boss, in this case, Satya Nadella, is telling investors that the company needs to win back its fans, clearly it understands that there's an issue here. So yeah, as of now in the release preview channel, you can now move the taskbar left, right, top, or bottom, you can make it smaller. You can adjust the Start menu as well, make that smaller or larger too. So it's all customization.
One of the concerns I've got is it's taken Microsoft five years to get to this point, which is a bit of a worry.
Avram (21:07) I give them zero credit for doing this. If you take something away from users and take five years to give it back, you don't deserve a pat on the back. They created a situation where they could congratulate themselves for restoring a removed feature. While I don't really care about having the taskbar at the top or the side, I'm happy to have it on the bottom of the screen But they took away a choice from me for no apparent reason. I know, it's the code. But you know what? That's not my problem. That's their problem. They wrote the code in such a way that they overrode users' choice. How long have you been able to move the taskbar?
Richard Speed (22:02) It goes back years and years.
Avram (22:03) Since Windows 95, hasn't it?
Richard Speed (22:05) I think there's two aspects to this. The first aspect is obviously the concern about how long it's taken. From my perspective, it's like there's been five wasted years. People complain about this, why is it taking so long to actually do this? Yeah, as you say, they don't deserve a pat on the back. They have wasted time before this putting in things that users perhaps weren't that concerned about, like the adverts, like Copilot everywhere. But it's also the company has recognized that there were certain things that it it did do right in Windows 10. A good example is the right-click context menu, which is now in the experimental channel. Now there Microsoft has actually come out and said, you can reuse your Windows 10 muscle memory here because we're going to allow you to customize it and make it work more like how you're used to. And I think that's really good they're doing that.
But I think it's it's a real shame it's taken them so long to actually do this. I'd recommend anyone who's got a sacrificial PC to have a look at the new context menu 'cause it actually works, I hate to say it, really well. They've done a good job, I think. It's just I wish they'd gone in that direction three or four years ago before inflicting their design choices on the public, which I think perhaps weren't that keen and they didn't seem to be able to listen to the reaction.
Tobias Mann (23:32) How much of this is motivated by the lack of competition when Windows 11 launched? Macs were pricey, and Linux was an enthusiasts' game and filled with compromise. Five years into Windows 11, that situation seems like it's changed quite a bit with, memory prices having driven up the cost of every PC to a level that, unless you have a large supply chain that can help you avoid some of that cost, everything's expensive. Macs are a lot more competitive price-wise than they used to be because Apple's supply chain is just better than most PC manufacturers'.
As well as on the Linux side of things, you know, there's Steam, which has basically made one of the one of the stalling points for Linux adoption with gaming a non-issue for the most part. And so it seems genuinely like Microsoft has to start investing in a platform that it thought was just kind of a monopoly.
Richard Speed (24:54) There are echoes of Internet Explorer, where Microsoft owned the internet browser space until Chrome came along and ate their lunch. I think what's happened is they've looked at the adoption of Windows 11 and it hasn't gone well. Even though Windows 10 is now out of support – you have to be on the ESU program to get any support at all for Windows 10 – people are still using it, and people still like it.
And I think what's happened at Microsoft in the last year is they've looked at this and thought, "Okay, we might have made a mistake here." Or the people in charge have refocused where they're going, thinking, "What's stopping people from moving from Windows 10 to Windows 11?" And yes, I think there's definitely more choice now than there used to be. Linux and macOS are good choices, and we're in this frankly bizarre situation now where Apple hardware is often the cheap option, which I still can't believe how we got here. But I think ultimately Microsoft is realizing that people liked how Windows 10 worked. Why did you change it? And so they're going back. My concern is, again, how long it's taken them to realize that. Windows is a supertanker – it takes a long time to turn that supertanker around. I think we've got another year or so before they get back to where they were, at which point they've lost four or five years of time, which is a real shame. But I'm happy to see what's happening.
Avram (26:40) So I do wonder how many things that are old are going to be new again. For example, the wallpaper. I know this is really petty, but these are the things people look at when they go to your desktop. The default wallpaper in Windows 11 is hideous. And I think people look at that and feel like you're even ashamed that it's Windows because it doesn't say Windows on it. It used to have the Windows logo; you used to be proud of your Windows. Now they have some weird papier-mâché flower that looks like it's wilted. They got rid of some of the special effects that people liked – people liked Windows 7 because it brought the Aero Start menu. So I wonder whether we're going to start to see a new version of Windows like Windows 12, and people will like it because it will just bring back some of the visual elements and choices that people had in 7 and 10.
Richard Speed (27:44) I think we'll certainly see more options to make this work more like it used to. In fact, today we saw an announcement in Microsoft's Message Center about Outlook, where you can make new Outlook look and feel a bit like Outlook Classic. It's a small thing, and it's cosmetic, but it's what people want. They prefer to have things work the way they're used to and happy with. There's no need to change things just for the sake of change.
Avram (28:21) Well that's why Windows 11 has been such a pain because when it started, it was just change for the sake of change. Now we've got new features like AI piled on top of it. Was there really any evidence that we needed to have the Start menu in the middle of the screen, or was that just "Look, let's show you that we can make a different move"?
Richard Speed (28:42) I think you're right. Things like rounded edges on window s –there were a lot of very cosmetic things with Windows 11. Over the last few years, I think Microsoft has come to realize that maybe things worked okay before and they didn't need to change this stuff. It just feels like a lot of wasted effort to get to this point.
Tobias Mann (29:02) Is this any different than what we saw in the late nineties with the transition from Mac OS 9 to OS X? That was so divisive that Apple had to ship both operating systems on those machines, and for the initial period was actually shipping Mac OS 9 as the default. Or the transition from Windows XP to Vista, which I recall as being heavily influenced by the stiffer hardware requirements, reminiscent of the TPU requirements on Windows 11. It seems like this isn't as dramatic a departure from Microsoft's playbook in the past; it just seems like maybe users have lost interest or are struggling to adapt.
Richard Speed (30:06) In this instance users probably have lost interest because Microsoft reached the point, certainly with Windows 10 and Office, where it did everything you needed it to do. So why would you need to upgrade? If I were cynical, I would say that with Windows 11 there was probably demand from hardware partners saying people need to buy new PCs. They were like, "Well, here's Windows 11."
"Windows 10 hardware works fine, why should we buy more hardware?"
"But you need to have this special security chip," and so users were then forced to buy new PCs or new devices. As you say, there wasn't a whole lot of difference, apart from the visual stuff, initially, between Windows 11 and Windows 10. Since then, there's been a lot more AI put into it for sure. But again, it's debatable whether there is a killer AI application on the desktop yet. Driverless cars are a great example of a really good AI application, but on the desktop, I'm not sure there is one yet. That is a problem Microsoft is also facing, with users saying, "There's nothing here that I absolutely have to have to do my day-to-day job."
Avram (31:22) I remember last year I wrote an article about the things I think Microsoft should actually do to improve Windows. I'm sure they won't do any of them, but I was thinking of things that would actually be good for power users. I guess that's my bias here. Having a second and third clipboard built into the OS would be a big deal. Maybe adding another modification key in addition to Windows, Control, Alt, and Shift, because all the keyboard shortcuts seem to be taken up by something. We could use another key combination for more keyboard shortcuts. I feel like there are things they could do to really make this operating system more friendly to an audience that knows what they want and is getting more power-user heavy, because people have been using Windows for 30 years now.
They don't need to make it simple; they need to make it powerful. And they're more interested in pushing ads and AI at you, then changing the UI just to make it look like they did something and get you to buy a new computer because they upgraded the hardware requirements. I feel like the reason people aren't switching to it is because they're annoyed. They're annoyed that they're being pushed to get new hardware that they don't want, they're annoyed that they're being pushed to deal with a slightly different UI when the old UI worked fine, and that they're having things like AI and ads pushed at them. But hey, speaking of AI being pushed at them, I understand that the new Task Manager is going to better keep track of when your AI is being used. Is that right?
Richard Speed (33:10) Yeah, this is another thing Microsoft pushed quite hard this week: Task Manager. I can remember when Task Manager first appeared back in the days of Windows NT, and it was all very exciting. Dave Plummer wrote it initially and called it a stone-cold killer because it was great – you could just kill processes. It did one thing really, really well: you could see what was going on and kill stuff. Whereas over the years, and you could take this as a metaphor for Windows, it's just gotten bigger and does more things, which maybe you don't want. The latest thing is you can get more visibility into AI workloads. So as well as seeing how hard your NPU (neural processing unit) is working, you can now drill down and see which processes are using it. The goal there is to see: "I've got a process here that should be using the NPU, but it's actually using the CPU and GPU, and therefore burning up much more power. It's far less efficient." So it does make sense from that point of view. However, as I was reviewing this, I thought, "Is this really what I want Task Manager to do?" I want it to be a very lightweight, simple tool where I can see what's running, what's stalled, and what's crashed, and kill it. That's all I want it to do. If I want to see what's happening internally, I want to use a tool like Performance Monitor or an inspection tool to see what's going on in my system. I'm torn about this because it's always good to see these tools being developed and being enhanced, but I'm not sure it's the correct place to put that process monitoring information. I want Task Manager to manage tasks and that's it.
Avram (35:00) I don't really have a problem with this personally.
Richard Speed (35:02)
(Laughs.)
Avram (35:02)
I like seeing what my CPU, GPU, and now NPU are doing, so put it in a place where it's easy for me to keep track of. That's fine, I don't have to look at it – I think that's a separate tab for the NPU, so you don't have to click on it. I don't see this as a huge problem as long as I'm able to see my tasks and kill my tasks. I want as much information as possible at my fingertips about what my computer is doing.
Speaking of things being at your fingertips, I want to thank both of you for coming on and talking about this week's important topics. As always, we will have lots of information at The Register at your fingertips, talking about Windows, AI hardware, security, and all the other things that we do to be the hand that bites IT.
We look forward to seeing you all next week. ®