Hands-On Windows 215 transcript

Please be advised that this transcript is AI-generated and may not be word-for-word. Time codes refer to the approximate times in the ad-free version of the show.

Paul Thurrott [00:00:00]:
Coming up next on Hands On Windows, we're going to take a look at a powerful new kind of AI computer called RTX Spark.

TWIT.tv [00:00:10]:
Podcasts you love.

TWIT.tv [00:00:11]:
From people you trust.

TWIT.tv [00:00:15]:
This is TWiT.

Paul Thurrott [00:00:19]:
Hello everybody and welcome back to Hands On Windows. I'm Paul Thurrott, and this week I'm going to begin an exploration of this new Windows PC platform we have called NVIDIA RTX Spark. So, and this is really hard to explain. I always want to kind of present this in a way that makes sense, position it wherever it belongs. And it's hard sometimes with this to understand where that is. But if you think about Windows computers over the past decade or so, we've had something that Intel and Microsoft called AI PCs. And those are PCs that have a dedicated NPU of any kind, right? The use case for that is really just one thing. It's Windows Studio Effects, which is mostly video effects, but a single audio effect.

Paul Thurrott [00:01:05]:
The number of effects and which effects you get will vary by hardware, but those early AI PCs were fairly underutilized and fairly underpowered from a local AI perspective. But that was the beginning. 2.5 years ago, we got something called our— sorry, Copilot plus PC. And this was a 40+ TOPS MPU, so really powerful local AI capabilities, lots of other requirements, and most importantly to me, lots of end-user features and a list of end-user features that improved over time, right? And these are things we've talked about: Click to Do, Recall, and then all the many features and all the many apps in Windows 11. You know, Paint and Photos especially have tons of AI-based image editing capabilities, Snipping Tool, Notepad, Settings, everything else. A lot of apps just have little individual features that are all made better by local AI. But now we have what I think of as the 3rd generation AI PC, which is the NVIDIA RTX Spark. And so this is a platform NVIDIA created.

Paul Thurrott [00:02:10]:
It has an ARM CPU and an NVIDIA Blackwell GPU like the ones they have in the data centers and also on in gaming rigs combined on a single SoC or system on a chip along with up to 128 gigs of unified memory. And unified memory matters because you can split it between the CPU and the GPU. There are ways to do that semi-automatically, but Windows doesn't yet have that kind of a scheduler. I think it will, but we'll, we'll get to that. Um, but in the meantime, you can, you can do that dynamically if, if you want. So the smallest system— not smallest from a resource perspective, not smallest in size, they're all pretty big— has 24 gigs of RAM. But the 2 most common configurations are 32 gigs and 128 gigs. Um, the 32-gig machines target creators, so these are people like who edit, you know, photo or video.

Paul Thurrott [00:03:08]:
Yeah, and things like that. I'm not really sure what else there is to say. I mean, there's not much there. But the other part of this market is even harder to describe because at 128 gigs, which is the most typical of the higher-end configs, I'm not really sure who those people are. They're high-end AI enthusiasts, they're data scientists, they're developers, they're whoever really wants to be in on the leading edge of local AI, which is a pretty small market. audience, I would imagine. Like I said, local AI is vague. I've got the system properties up here just to kind of show you what this looks like because this is, it's just kind of a unique configuration.

Paul Thurrott [00:03:48]:
So this particular computer has 128 gigs of RAM, which it accurately says here. Half of it roughly, it says 62, 62.5 gigs is dedicated to graphics. It's really half, but we'll just call it 62, 62, 64, whatever between GPU and CPU. It's kind split evenly. That's the, that's the default. We'll look at how to change that in a second. Um, and then whatever storage. And then there are 2 versions of this.

Paul Thurrott [00:04:16]:
Uh, it's an SoC really. So, um, it's called NVIDIA RTX Spark N1x. Um, this— it's— this higher-end version is a 20-core CPU and a 6,144-core GPU. Um, the lower-end version, I think it's an 18-core CPU and probably a 5120 or somewhere in there GPU. The cores in the GPU are not all created equal, right? So one of the things that they do is they have different types of cores and some of them are super low power, super low, you know, high efficiency, whatever. They combine, I think it's 2 or 4 of those cores to create what is essentially a kind of a software-based MPU. So it appears as an MPU to the system. We'll probably look at this few times, but if you go into Task Manager, um, keep clicking the wrong thing, you can see it says it has an NPU.

Paul Thurrott [00:05:06]:
And it's, uh, I don't know if it's in here, but somewhere it's identified as an NVIDIA NPU. And NVIDIA doesn't actually make an NPU, you know. In fact, everything I'm going to do on this, and most people are going to do this, is going to hit the GPU. But it does have this Blackwell GPU, which is awesome. So let's see what makes this thing different from what came before. Well, if you think back to Copilot+ PC, initially it launched on ARM with Qualcomm chips, so Snapdragon X and X2, but Intel and AMD both came up with versions of their chipsets and then computers based on them that are Copilot+ PCs as well, right? And so somewhere between a 40 and an 85 TOPS NPU, 16 gigs of RAM. I think the minimum on storage is 256, but usually it's a little bit more than that. And all of those things get all of those features, you know, Click to Do, etc.

Paul Thurrott [00:06:01]:
This thing though is curious because it's not just an ARM-only architecture in the sense that it runs Windows 11 on ARM, but it also has this NVIDIA GPU, which has never been on ARM before. Right. And like I said, the CPU is kind of made up of whatever cores. The reason it's— this architecture exists is for the unified RAM bit, and it can be split between those 2 processors. This is something Apple gets automatically by the virtue of their architecture. So they started out with phones with the A-series chips on Apple Silicon. They have M-series chips now on Macs and other chipsets. But when you kind of move forward to the Mac and with this chipset they have, they have unified memory and they can actually scale higher than we do even with this brand new platform.

Paul Thurrott [00:06:47]:
And so This is the highest-end config that's possible. It does not scale beyond 128. But if you spend enough money, you could get a, you know, Mac Studio or maybe a Mac Mini with whatever the latest and fastest, best Apple Silicon chip is, probably M5 Pro Max or Ultra or whatever, up to 512GB of storage, which is far more than this. And also it can be faster. So at the M5 level, the memory speed is just about the same. But once you get up to M5 Pro, M5 Pro Max, or Ultra, whatever the names of those chips are, it gets dramatically faster. I mean, like up to 4 times faster. It's a big advantage that they have over there.

Paul Thurrott [00:07:28]:
Now, Nvidia is very popular with software makers, of course. So, you know, we're probably going to see things that are optimized for the system. And down the road, I'm sure we'll have faster systems and more RAM as well. Also very expensive, I should say. These computers run $3,000 to $7,000 depending on the configuration. So the base configs are usually right around $3,000. $3,000 isn't bad compared to some personal workstations or portable workstations or really high-end premium laptops. $7,000 is more than I've spent on most cars.

Paul Thurrott [00:08:03]:
I don't, you know, I'm not sure what to say to that. But to me, the big— the biggest difference and really the biggest issue is that where Copilot+ PC had specific end-user features out of the box, you just got them for having the computer, RTX Spark does not. You really have to know what you're doing to use this thing. So this is— this desktop is— this is RTX Spark. And so I already showed you this, but the System About area here in Settings will show you that this is the highest-end configuration, etc., etc. One of the oddities here, and I don't, you don't see it here, I guess, but if we just run WinVer, you can see it here, is that this thing runs 26H1. So this is the version of the operating system that shipped in the spring when Qualcomm shipped the Snapdragon X2 series of microprocessors, which are also Windows 11 on ARM. And at that time, Microsoft and Qualcomm were saying that this is just because of some unique need of the silicon.

Paul Thurrott [00:09:01]:
It's not going to be upgradeable to 26H2, but somewhere down the road we're going to talk about where you can go with this thing next. In other words, there'll be some future version of Windows, whatever it might be, 11, might be 12, who knows, that will be an upgrade for these things. But since that happened, 26H2 has come out, 24H2 has gone away. If you were on an earlier version of Windows other than 26H1, you will have either upgraded to 26H2, or you're not able to upgrade for whatever reason, but these things remain on 26H1. And then NVIDIA Spark, RTX Spark-based computers came out, and they're also on 26H1. So I don't actually think this is tied to the Arm architecture, but perhaps I'm wrong. But for some reason, these 2026-era new silicon releases are on this kind of weird, unique version of Windows 11. It doesn't matter too much other than the fact that If you hear about new features coming to Windows 11, this version is about a month behind.

Paul Thurrott [00:10:01]:
So if Microsoft adds a new feature to Windows 11 in, say, October of whatever year, it won't come here on 26H1 until November typically. I don't know why. It's never made sense to me. As far as the whole software experience goes, it's Windows, right? So we have a Start button, Start menu works as before. The taskbar, even though it's a month behind, it does have those new capabilities. So I can position it on different sides. I'm using the small taskbar, which I love. Windows Search does not have the search highlights, so that's been cleaned up nicely, which is great.

Paul Thurrott [00:10:39]:
You know, widgets, etc. All these UIs exactly the same. I mean, just from a day-to-day perspective, there's nothing that stands out. You know, one of the little oddities is that it's supposed to be RTX Spark. A superset of Copilot plus PC, and it does have that NPU, like I said, the software NPU. But if you look at the Copilot menu here in Photos, and you'll see this in other apps too, you actually, right now at least, I expect this to change just based on the way Microsoft's described this, but it doesn't have those local AI features. So not only does it not have like an obvious starting point, it doesn't even have the things that came in Copilot plus PC, which doesn't make a lot of sense to me, but That's where we're at. I had showed you in the Settings app what it looked like from a kind of a RAM— yeah, well, like an allocated RAM perspective, which was showing you all the RAM, right? And if we bring up Task Manager again, you can see here it believes that there's 64 gigabytes of memory, which is really interesting because this is actually more accurate.

Paul Thurrott [00:11:41]:
So there is in fact 120 gigs of RAM, but it's split evenly between the CPU and the GPU. So what it's showing here is the amount of RAM that's available to other running— like apps that are not running on the GPU, essentially, meaning most apps, right? But you can configure how that works, just not in Windows. And so the way you do that, oddly, is through the NVIDIA app, which is kind of bizarre, right? If you have ever played games on a PC and if you've had any NVIDIA graphics of any kind, whether it's on a laptop or a desktop, you've probably dealt with this app or its predecessor in the past. Most people hate it. I do too, but they don't force you to log in anymore, which I like. And there's a lot of stuff in here now that's kind of specific to this platform, which is really interesting. So as before, you will, you know, they'll prompt you, but you can come in here to figure out if you're up to date with the latest drivers, which you absolutely want to be on the system. There's a couple of different choices for driver types.

Paul Thurrott [00:12:38]:
Obviously, if you're going to be playing games, which is not necessarily why you're on this platform exactly, although we'll get to games in a bit. You could use that. That actually is what I'm using because I actually am playing games, but we'll get to there. But if you want more stable drivers and you're a creator, developer, whatever, you can choose this version instead. So I've got— I'm stuck with that. It seems to be fine. And then a couple of different things as we kind of— I went to that already, I'm sorry— as we dive in here. So I don't have any games on this particular laptop installed, so I have 3 of these in-house right now, and so I've, I've games on other computers.

Paul Thurrott [00:13:17]:
So what I had to do was just take some, bring some screenshots over so I can at least show you what this might look like elsewhere. So this is the, this is what this screen, so in other words, this screen from the NVIDIA app on this computer looks like on one of the other computers. And so this This particular computer, I believe, has 32 gigs of RAM total, and it has some games installed. And of— there's actually only 2. I don't know why it has 3 listed there. But, um, Gears of War is an example. It's a brand new game. It's optimized for this platform, and it runs great.

Paul Thurrott [00:13:53]:
And I, I actually tried to install it on here before I started the show, just so I could kind of show it to you. But, um, I do have— I have some screenshots or whatever. But, um, It's run surprisingly well. So if you're familiar with gaming on ARM, you know that it's not as good. Like Snapdragon X2 is a great chipset, but it has to emulate games. The way Nvidia describes this is that because these games run on the GPU, part of them actually runs natively. And so you get much higher performance, better graphics quality, and a wider range of games that are compatible out of the box. Like next year, for example, the new Call of Duty will support this platform.

Paul Thurrott [00:14:27]:
right, which is amazing. Gears of War works, Doom: The Dark Ages works great, the latest Halo game works awesome. I've installed games from other stores like The Last of Us Part 1, which I think I got from either Epic or Steam, I don't remember, works wonderfully as well. And I think, yeah, this is probably the— oh, that's the screen I already showed you. Nope, that's not what I wanted. Let me just Let's do it this way. So, you know, I just ran— I just, as I played the game, the original, whenever that was, a month ago or whatever, 2 weeks ago, I just kind of stepped through it. It does that shader loading thing, which everyone loves.

Paul Thurrott [00:15:04]:
And then you can kind of see how it does. And, you know, again, for a RAM system— RAM system. An ARM system, this is rather incredible. Not quite 60 frames a second, but effectively 60 frames a second. Really nice performance. You know, and it looks great. I mean, these are just— these are non-HDR screenshots, but, um, you know, it looks really nice as well. And that's a different game, but you get the idea.

Paul Thurrott [00:15:26]:
And then I also ran some benchmarks on Doom: The Dark Ages, and there are 3 different benchmarks, but effectively speaking, this thing is over 60 frames a second. It's running native res, which is amazing. It's running at high graphics detail. It just looks awesome. So that is I mean, that's something very, very different. That's really cool. Um, and then what else do we got here? So there's a couple other things in the NVIDIA apps. Such a weird thing to use the NVIDIA app as kind of a core system utility, but you know, that's where we're at.

Paul Thurrott [00:15:57]:
Um, you can, you can set some settings in here as well. Um, you can go into System, and this is going to have, uh, 2 new things if it ever loads. Um, in Video, you can turn on Super Resolution for video, which I've done. And so let's see, I think I can go to Edge and just go to YouTube. I think any video would probably work. And this is sort of like— well, not sort of like, this is just like Auto SR for games in Windows 11, especially Windows 11 on ARM, right? In other words, it can run this video at a lower frame rate or lower quality level, but the effective frame rate/quality level to you will be very, very good. And that saves bandwidth and it You know, it runs better, et cetera, et cetera. So I don't know why I'm looking for a particular video here, but okay.

Paul Thurrott [00:16:46]:
So Apple finally has touch laptops. Bring this guy up.

TWIT.tv [00:16:49]:
If you give a Mac an OLED, it'll have to be a different kind of—

Paul Thurrott [00:16:52]:
I will pause that so you don't have to listen to her. Um, but eventually I'm not seeing it maybe because it's not OLED,

TWIT.tv [00:17:00]:
like a Pro.

Paul Thurrott [00:17:01]:
So what this is showing you is that it is using, uh, NVIDIA GPU technology to Smooth out the frame rate, improve the graphics quality, lower the bandwidth, etc. So that's, that's what I enabled by just checking this box here in here. But bigger news occurs here. So there's 2 things. Dedicated GPU memory— this is where you configure how much the RAM goes to CPU or the GPU. This is the default on the system. It says 62.5, but essentially 64. You can't actually go up to 128 because then there would be no RAM left for the system, right? But I think it goes up to about 92, 94 gigs, something somewhere in that category.

Paul Thurrott [00:17:41]:
If you make a change to this, you have to reboot the computer for that to take effect. So it's not something you're going to do on the fly. This one has appeared fairly recently, so this wasn't here when I initially got these laptops, but now you can schedule app prioritization. So, uh, it really— it's GPU prioritization, I should say, um, whether for foreground or background apps or just the default, which is split evenly. Depending on how you use this computer, you might want to pay attention to this, right? So it doesn't make sense for a laptop, but if you think about an RTX Spark— there are desktop versions too— whatever this box is, as something where you're going to run AI agents and have it be there all the time, you'd probably want this thing to be heavily tilted toward background app because you're not going to— you're going to be using it headless essentially. But a laptop, it's a little more complicated. You're interacting with it. So I'm not 100% sure where to go with this one, but this is brand new.

Paul Thurrott [00:18:29]:
So it is— it's kind of an interesting, um, kind of an interesting new development there. And then I think there might be one more thing in here. Uh, that's already on by default, so that's fine. Okay, so why does this exist? That's a good question. Uh, I've been struggling with that one for a couple weeks, honestly. Um, but the reason this exists is Local AI, and more specifically local AI agents, right? In the same sense that we've had 3 generations of AI PCs, in some ways we've had 3 or some number of generations of just AI in general. And then this local AI thing mirroring it, sitting off to the side, not as capable, more complicated, but free, something you can put on your computer, you can run privately, you can run offline. There are, you know, obviously there are advantages to it.

Paul Thurrott [00:19:24]:
But no actual end-user features to take advantage of any of that. So you're on your own. You get this, you're presumed to be an expert. I consider myself fairly technical, and I have to say I've struggled with this because there are just so many choices to make and where to look or what the right answers are is not always obvious. But we're going to look at that next week. So next week on the next episode, we're going to take a look, take our first steps, I guess, because it's probably a lot to this, um, of what local AI looks like on these RTX Spark computers. So we'll do that next. Alrighty, well, thank you for watching.

Paul Thurrott [00:20:05]:
Hopefully this was useful in some way. Um, we'll have a new episode of Hands On Windows every Thursday. You can find out more about this podcast at twit.tv/how. Thank you for watching as always. Thank you especially to our Club Twit members. We love you guys.