Nvidia DGX Spark as a Mini Desktop PC? - Featuring Dell Pro Max GB10

For the last month, I have been playing around with this, the Dell Promax GB10 mini workstation. But don't let its small size fool you, as this is rocking a 20 core processor, RTX 570 class graphics, and a full 128 GB of unified memory. But while this is an amazing piece of kit, I'm not sure if everyone's going to run out and buy one for themselves. Let's talk about it. Today's video is brought to you by MicroEnter, your professional AI destination. Whether you're an AI dev or an entry-level enthusiast, MicroEnter is the premier destination for AI hardware. You can fine-tune your own models with the latest hardware from Nvidia, like the DGX Spark AI workstation, or an RTX Pro 6000 Blackwell graphics card. How about the latest gear from AMD, including Stricks Halo powered desktops and laptops, featuring up to 128 GB of shared memory? Of course, MicroEnter is still your source for gaming PCs, laptops, and components as well, and have everything you need for your next builder upgrade. MicroEnter is opening their 31st location in Austin, Texas. And if you follow the link down in the video description, you can snag a free 128 GB flash drive for yourself when they open later this year. Also, make sure to head over to microcenter.com and sign up for MicroEnter news for the latest on PC hardware. Thanks again to MicroEnter for supporting Craft Computing. And now, on to the video. Welcome back to Craft Computing everyone. As always, I'm Jeff. Before we get started, a quick standard disclosure. Dell did send out the GB10 as well as their brand new U5226KW monitor for me to take a look at, and both will be returning to Dell once I've had my fun with them. No money changed hands. Dell gets no input into the production of this video, nor will they see it before it's released on YouTube. At CES 2025, Nvidia dropped a bit of a bombshell with a tiny PC known at the time as Project Digits. Fast forward to their annual GTC conference and the gold box got an official name in the Nvidia DGX Spark. The Spark is notable as Nvidia has often shied away from releasing 100% in-house platforms, let alone fullyfledged PCs. We've certainly seen some devices that are powered exclusive by Nvidia architectures before. The Nvidia Shield and the Nintendo Switch and Switch 2 both come to mind, but those were more or less purpose-built devices for gaming. The Spark is a generalpurpose PC running standard Ubuntu Linux, meaning you can use the Spark to do whatever you want it to do. There have already been a number of reviews on the DGX Spark or the GB10 showing off Nvidia's intended use case, that being a local AI server or development workstation, and I'm going to link some of my favorites down in the video description. Shout out especially to Jeff Gearling for his awesome work and comparison between a pair of Dell GB10s and a four node Mac Studio cluster. That one is definitely worth the watch. Now, I'm no stranger to AI, having done a couple of recent videos benchmarking AI workflows and inference. But to be honest, the more we're learning about the capabilities of current generation self-hosted AI, we're also seeing more and more of the limitations of LLMs in general. While highly trained and specialized LLMs can be incredibly powerful tools, generalized models tend to hallucinate more often than not. A problem that's actually getting worse, not better. And that brings us to today's video on the table. Next to me is not the Nvidia DGX Spark, but it is essentially the same exact thing with the same internals. This is the Dell Pro Max with GB10, a dimminitively sized desktop PC with some seriously impressive specs. At the heart of this PC is Nvidia's GB10 APU, and things get weird right from the jump. Rather than relying on Team Blue or Team Red to supply CPUs, Nvidia is using a semi-custom ARM CPU to power their vision of a mini workstation. The CPU section consists of 20 cores in total in a big little arrangement. 10 ARM Cortex X925s for high performance and 10 ARM Cortex X725s for lower power workloads. And as far as the GPU, well, that's what Nvidia does best with what's essentially an RTX 5070 on board. We're looking at 6,144 CUDA cores, 384 tensor cores, and 48 RT cores, albeit with some limitations when it comes to memory. As this is an APU, the memory is shared between the CPU and GPU dynamically as needed. Given that there's a 128 gigabytes of LPDDR5X, there's plenty of memory to go around no matter what you intend to do with the GB10. Be it an AI workstation or just a super compact daily driver. But like I mentioned, memory bandwidth is severely limited compared to its desktop counterpart. The RTX 570 desktop card has a 192-bit memory bus, but is built with GDDR7, giving us a total memory bandwidth of around 672 GB per second. While the GB10 has a wider 256-bit bus, it's relying on a shared pool of LPDDR5X. Thus, the GB10 is limited to a max of around 273 GB per second, and that's without the CPU stealing any of its resources. The rest of the GB10 specs are pretty standard fair. There's a single HDMI 2.1B port for video out, four USB type-C gen 2x2 20 gigbit ports for all of your IO needs. So, make sure to bring a dongle or a dock of some kind if you still have USBA connected keyboard and mice. There's integrated Wi-Fi and Bluetooth thanks to a MediaTek 7925 Wi-Fi 7 controller. And if that's not enough, there's also a RealTech 812710 GB nick with an RJ45 jack around the rear. And if that's still not enough network bandwidth, how about a pair of connect 7 200 gigabit nicks each with their own QSFP plus port. While the intention here is for the ConnectX cards to be used to cluster multiple GB10s together, it is just standard Ethernet. Meaning if you have 100 or 200 GB networking available, you can take advantage of it with this tiny little PC. While buying a couple of Dell GB10s, stacking them together, and running local AI training for small data sets or inference on 100 billion plus parameter LLMs is pretty cool, that's not what I'm doing in today's video. See, between Nvidia, AMD, Intel, and Apple, we've gotten some amazing APUs over the last couple years. CPUGPU combos with shared memory are no longer the solution that's only slightly better than integrated graphics and still only 10% as fast as a dedicated GPU. Modern APUs are actually rocking desktop level graphics and often for quite a bit less overhead than you might expect. That graphics power is why I am so excited about systems like this. The Dell GB10 is easily able to drive massive displays like the Dell U5226KW over my shoulder here with its 6144x 2560 resolution and 120 Hz refresh rate. And with four 20 gigabit USBC ports, connecting multiple displays or even daisy chaining them together should be no problem at all. There is one massive downside to the connectivity, though. See, the USBC ports are only Gen 2x2 and support only display port 1.4, which means it's limited to 60 Hz on panels with 4K or higher resolution. While the monitor is beautiful, it takes up the only HDMI port on the PC. And while the Dell GB10 graphics are more than enough to drive four or more of these displays, Nvidia clearly intends their Spark-based systems to be AI only boxes, not productivity machines or standard desktops, which is a bit of a shame. Like I mentioned, the Dell GB10 is just running Ubuntu Linux 2404 LTS, and firing it up for the first time drops you onto a standard Ubuntu desktop, and most of the process of using Linux on this machine is pretty straightforward. their standard apps like Firefox, which all just work as they're natively compiled for ARMbased systems like this. But there is a downside. See, the Ubuntu app store unfortunately doesn't seem to be aware of the hardware that you're running on a particular PC and recommending apps that are native for your architecture. All the apps are just listed alongside native x86 apps that likely are not going to work when installed. Not only that, but when I went back to the Ubuntu software store to record what I was seeing, the store just straight up stopped displaying any apps at all. In the first couple weeks of using and testing out the GP10, I could see the apps, but now the whole store just seems to be broken on this particular install. But like I said, because we're not running an x8664-based CPU, software compatibility is going to be very hit or miss. I attempted to install quite a few different applications just to be met with a series of impassible roadblocks. Be they outof-date dependencies, missing drivers for ARMbased hardware, or any number of other problems. I wanted to see if Da Vinci Resolve would function only to be met with basically every issue possible on this platform. Lack of up-to-date packages for half a dozen dependencies and ample warnings about complete lack of GPU acceleration features, even if I got it to run in the first place. So, for software, your mileage is definitely going to vary. But what about the Oblivion remaster loop that's been running over my shoulder this entire video? See, a couple of months ago, Valve announced one particular piece of hardware that many people are excited about, but I got excited about it for some very different reasons. The Steam Frame is a self-contained VR headset, complete with both inside out tracking for VR games, compatibility with both x86 and Androidbased apps, or the ability to stream from your gaming PC, all of it powered by an ARMbased APU. It's that last bit that really stuck out to me, as Valve is promising compatibility with not only Windows-based games, but the ability to translate them from x86 to ARM architecture to be able to do it. A small footnote during the announcement was a massive amount of contributions by Valve into a little piece of software called FEX. While not available by default yet, there are a couple of different installers available that will wrap the x86 Linux version of Steam into FEX, essentially allowing you to install any x86 game and play it at near native speeds, very similar to how Proton allows Linux compatibility with Windows executables and APIs. While we know a lot about the graphics power inside of here being roughly on par with an RTX 570, the CPU doesn't have a lot of comparison benchmarks to compare with hardware that we already know. So, the first thing that I did was one of the most cursed benchmark setups I've ever completed, and that is running Cinebench R15 through Steam powered by FEX. Shockingly enough, this worked remarkably well, giving us a multi-threaded score of 3870 and a single threaded score of 275 with 20 cores. Those marks put it in essentially a dead heat with an Intel 13600 K, which is also a 20threaded chip. AMD's stricks Halo AI Max Plus 395 with its 16 cores and 32 threads of Zen 5 are obviously faster here, coming in at scores of 5886 and 317 points, respectively. The CPU Instrics Halo definitely has the upper hand, while Nvidia's GB10 put more of its eggs into graphics performance, but honestly, not a terrible showing for Nvidia in their first semi-in-house CPU offering. And since we're in Steam, we might as well look at what ARM compatibility is looking like in games, seeing as Valve's Steam frame is due to launch in the next couple of months. Honestly, this reminded me of my first look at Valve's Proton compatibility layer back in 2019, as it's pretty incredible how much they've both got working and got right in this very early preview. Most games that I tried just worked. And with a mid-range CPU with an RTX 570 class graphics processor on board, performance was downright pretty solid most of the time. Things are definitely slower than I'd expect on the CPU side of things with games like Cyberpunk failing to hit 60 fps on average at 1080p and ultra settings. The jump to 4K only dropped me down to about 45 fps as well. So, there's definitely some CPU performance that's holding us back. Unfortunately, the Steam overlay only worked in about one out of every 20 games that I tried. So, I really couldn't get much data beyond in-game FPS counters. So, don't consider this a full official benchmark. I'm sure getting it working is pretty high on Valve's list though as it's a pretty vital component of Steam and coming to the Steam frame as I mentioned in a couple of months. Kingdom Come Deliverance 2 and the Oblivion remaster were both basically flawless hitting around 45 fps at 4K and closer to 60 fps at 1080p. Oblivion during indoor scenes was actually closer to 90 fps, but struggled in some outdoor scenes, which is pretty common for this title. However, over my shoulder right now is Oblivion running at ultra settings with ray tracing enabled, and it's doing about 55 fps outside of an Oblivion gate. So, this is what it looks like. Starfield was also a pretty decent experience as well, managing around 60 frames per second at 4K and ultra settings. Doom Eternal, Stardew Valley, Tearown, and quite a few other games also all ran without any issues. Crab Champions has been one of my favorite games for the past few months and is nearly impossible to accurately benchmark as the scale the game can build to is completely off the charts. Early Islands will spawn 30 enemies and you'll fire one shot every couple of seconds. By Island 28, you're shooting around 10 shots per second at 250 enemies. By Island 70, it's a bullet hell game the likes of which you've never seen. But every projectile you fire does damage and multipliers and bounces and different status effects. All of which has to be tracked. And all of that noise does very much bog down your CPU, especially in later levels. As such, performance is pretty mixed from early to late game and hell on both your CPU and GPU. In the case of the GB10, though, I found an interesting glitch. If you stress out the CPU to a point where you drop frames, the game will flat out never recover, even when all the assets have been cleared and the maps are reloaded, your average will go back to 120 fps, but you seem to drop every second or third frame, making the whole game stutter like mad. What does this mean for the GB10 right now? Not much. As remember, we're playing a game that was written for Windows, running Direct X11 for graphics, and x86 instructions for your CPU. We're running the game in Linux by translating DirectX calls into Vulcan and translating x86 calls into ARM instructions. A couple glitches are to be expected, especially this early on with fax. But like with Proton, I expect nothing but dramatic improvements over the next couple of years. Games that I didn't have a great time with included Red Dead Redemption, where I was technically able to launch the game and get into the settings menu, but the game crashed upon loading the single player mode. Hell Divers 2 launched and seems to run at a decent speed. just uh buy some sunglasses if you want to see anything. Circling back for a minute, with Steam getting so good at running Windows games on Linux and now even able to run on ARM processors, I decided to try and install the Windows version of Da Vinci Resolve through Steam to see if I could take advantage of its compatibility features. While I was able to launch the installer and get at least the startup screen to launch, Da Vinci would crash before even reaching the new project prompt, let alone getting into the editor. With all of the improvements that Valve has made to games compatibility, it's going to take an equal effort from more parties to bring those same functions to standalone Windows software. It's definitely a future that I hope for where the OS and hardware that you run on your desktop is pretty agnostic as any OS, any hardware, and any software can run on any hardware that you would like it to run on. That sounds pretty pie in the sky, but Valve is 90% of the way there with their games library translating from Windows to Linux and not nearly as far-fetched sounding as it was 5 years ago. So then, the Dell ProMax GB10, Nvidia's first consumerbound in-house PC platform. What's the verdict? Nvidia is obviously hoping to sell these as AI development and inference machines to businesses and entrepreneurs, as AI makes the bubble go up, right? I tend to view this with a bit more of a holistic lens and what Nvidia entering the PC market might mean for us as consumers in the near future. Apple is over here in one corner with their ARMPowered APUs and their own in-house hardware and software division. AMD is in another with their domination in desktop, laptop, gaming handhelds and game consoles market. Intel is in another with upand cominging moves to expand their GPU business along with Panther Lake and future desktop chips looking like they're having a very promising future. And then Nvidia might be the biggest player in the PC and data center space right now, but it's all of the other companies that are building out full systems with 100% in-house hardware. Nvidia is the only one that really isn't doing that, especially on the consumer side. When consumers have real choice and companies are competing for market share, mind share, and integration deals for other companies, we as consumers win. The AI bubble certainly isn't going to be around forever. And as cool as they are, a half stack of NVIDIA B300s is unlikely to land in my home server rack anytime soon. The Nvidia GB10 APU to me is a product that is serving only one market today, AI enthusiasts and early adopters at that. but it might have some massive implications in the near future when AI does indeed tumble back down to Earth. Nvidia now has a CPU product that they can tie to their GPU business that is consumerf facing and OS agnostic, something they didn't manage to do with the Nvidia Shield or the Nintendo Switch. And the more choices that we have as consumers, the better hardware we get to choose from. But there is one last thing we haven't talked about, and that's the price. The GB10 is expensive. Heck, it was expensive before the RAM and storage shortage, with Nvidia commanding a $4,000 price tag for their DGX Spark. The Dell Pro Max GB10 with a two TB Gen 4 SSD tips the scales at an eyewatering $4,200. That is double the price of an AMD Stricks Halo machine with a Ryzen AI Plus 395, a CPU that dramatically outperforms the 20 core big little ARM CPU Nvidia opted for in the Spark. However, the graphics on the Spark are also significantly faster than AMD's stricks Halo. So, six one, half a dozen another, I guess. But at the end of the day, a 20 core ARMbased CPU that runs standard Linux yet has no real software support beyond fledgling projects like Valve or natively compiled ARM ports of desktop software. A CPU that's only on par with an Intel 13600 K, even when you pair it with an RTX 570 class graphics card, just isn't worth four grand. Even when you weigh in the dual connect X7 network cards and the 128 gigs of memory, as an AI product, sure, you can name your price and if you slap AI onto something, you can charge 4 grand for it, right? But as a desktop product, I would love to see machines like the GB10, like the DGX Spark, and like AMD Stricks Halo down closer to the $1,300 price range. Configurations of 64 or even 32 GB of memory would do quite well at that price for use as a solid desktop PC. But the AI craze must go on. So to hell with affordability. Just trust us that this PC is a fully agentic AI machine and you'll never have to interact with it again because it'll do all of your programming and game development. It'll answer your Zoom calls and take notes and then take actions on the notes that that Zoom call takes. Right? But at this point in the AI craze, we know those claims are especially in the very near future. But until AI sucks the water dry, the only products we're going to see in the PC space for the next couple years are going to be AI powered snake oil. A solution in search of a problem sold at highway robbery prices, which is a shame because the GB10 is pretty freaking cool. While I'm down on its intended use, the GB10 is indeed a very novel system, and I really am excited for Nvidia to enter the consumer PC space. But being priced as an AI appliance means adoption is going to be limited to those with pockets deep enough to set up an in-house AI server only to be decommissioned less than a year later because none of the AI workflows wound up actually improving anything for your business at all. Here's hoping hardware like this isn't just a flash in the pan for Nvidia and we get more varied and more affordable options in the future because I really like what they managed to put together here. So yeah, I know not the most flattering of reviews, at least as far as a desktop use case and Nvidia entering the consumer space goes. As far as an AI product, I'm sure it's got pretty decent performance. And if that's what you're after, again, I have some reviews for those down in the video description. If you are interested in picking up the Dell Pro Max GB10 for yourself, I will also have affiliate links down in the video description. Make sure to go give those a look. On your way down there, make sure to drop this video a like and subscribe to Craftcomputing if you haven't done so already. And if you like the content you see on this channel and want to help support me in what I do, go on over to patreon.com/craftcomputing and get signed up. And that's going to do it for me in this one. Thank you all so much for watching. And as always, I will see you in the next video. Cheers, guys. Although maybe not with this one. Beer for today is one that I've been looking forward to for a little while now. This is uh sent in from good friend of the show, Aaron. Uh this is from Alvarado Street Brewery. It is the Caribbean Passion Imperial Sour Ale with passion fruit, mango, peach, strawberry, and orange. And it clocks in at 10%. What's more is this is not the new age jammy sour. This is a traditional sour ale. So for those who don't like drinking pie out of their glass, this one might be for you. May not be a jammy sour, but man is that thing still thick. Oh my god, I'm drooling. Do you hear sour ale? And do you hear 10%. And those are words that make sense in your head. They no longer make sense once you put that beer into your mouth. I understand it now. Holy crap. That was an involuntary shiver. Wow, that is sour. All right, I'm starting to get some flavors out now that the shock is gone to my tongue. I'm getting peach and mango more than anything else. Uh, and unfortunately those are two of my least favorite fruit flavors. I'm not a huge mango fan and I'm definitely not a huge peach fan. I was looking forward to passion fruit strawberry orange alla the orange Julius style of of uh Imperial Sour here. But the peach and mango, those are up front and center unfortunately. If you're into that, cool. Uh I was hoping the other flavors might shine a little bit more. So, at this point, I'm about halfway through shooting this video, and I'm this far into the beer. I'm trying I'm I'm trying to I was really excited for this one, and I always approach beers with an open mind. Um, I'm struggling with this one, though. That instant pucker factor is very real. Uh, this is a sour sour beverage. This is beyond Warhead candies. This is beyond just about anything that I enjoy eating. I'm not exaggerating that initial touch of sour. As soon as it hits your mouth, you are instantly puckered beyond what is comfortable. Uh, on top of that, the acidity is causing every salivary gland in my mouth to just go into overdrive. So, I'm shrinking my mouth, tightening every muscle, and also producing enough spit where it feels like I'm literally drowning. Positive things to say. I will say even though I am not a fan of peach and mango and those are the predominant flavors that are standing out to me in this beer, they're kind of nice. I I I think if you're a fan of those flavors, you would enjoy this beer. As long as you also enjoy hellaciously sour things. Uh I don't know that this one is for me though. In fact, sorry, I do know this one is not for me. Not every beer's a winner and not everything is made for me. And that's perfectly okay.

Thanks to Micro Center for sponsoring today's video. Micro Center is your Pro AI Destination: https://bit.ly/4aTcNim Check out their latest Pro AI Gear: https://bit.ly/4aECrGw Register to receive your FREE 128GB Flash Drive when their newest store in Austin, TX opens later this year: https://bit.ly/4tIr5K0 And don't forget to sign up for Micro Center News for the latest in PC Hardware: https://bit.ly/4qIFKCi Grab yourself a Pint Glass at https://craftcomputing.store Follow me on Bluesky @CraftComputing.bsky.social The DGX Spark is Nvidia's vision of an AI accelerated future, where a tiny appliance can sit on a shelf and power your entire home or business with agentic workflows and automations. But I'm more interested in it as a Workstation and ARM-powered desktop... so today, we're going to explore running software and games on the Dell Pro Max GB10. But first... What am I drinking??? Alvarado Street Brewing (Salinas, CA) Caribbean Passion Imperial Sour Ale (10.0%) Check out @JeffGeerling's review of the GB10 for AI Workloads and Clustering: https://youtu.be/FjRKvKC4ntw Check out the Dell Pro Max with Nvidia GB10 here: https://www.dell.com/en-us/shop/desktop-computers/dell-pro-max-with-gb10/spd/dell-pro-max-fcm1253-micro Dell U5226KW 52-Inch 6144x2560 120Hz Ultrawide: https://www.dell.com/en-us/shop/dell-ultrasharp-52-thunderbolt-hub-monitor-u5226kw/apd/210-bthw/monitors-monitor-accessories Support me on Patreon and get access to my exclusive Discord server. Chat with myself and the other hosts on Talking Heads all week long. https://www.patreon.com/CraftComputing