I would be surprised if it didn’t.
Probably not officially though, because of whatever agreement with Microsoft.
I think because Windows can only run on some of Qualcomm’s ARM chips, it would not easily run on the GB10, so nvidia created a new version of it that would more easily support windows. They probably also wanted to reduce the power consumption, so there could be a version for laptops.
I also have to think nvidia has planned all along for consumer applications of the GB10 (like when spending on data centers starts to level off or decline) - as I, like a non-trivial number of DGX customers, did not buy them for the same purpose most people on this forum did. I wanted to run AI locally, and a lot of people here have contributed a lot of knowledge to nvidia’s effort to make consumer hosted inference viable (it would be way more viable if the AI data centers didn’t consume so much DRAM).
I hope the systems based on the N1 and N1X will support linux from the start, I haven’t used windows regularly in over 10 years and don’t want to go back to it for an inference-capable laptop. I also hope these systems are not as late as the DGX Spark.
@VincentKaufmannUAE if the RTX Spark will use ACPI and not DT then any Linux distro will work. The DGX Spark uses ACPI but for the RTX Spark there’s no confirmation yet.
@elsaco because RTX Spark is just DGX Spark without the CX-7, and with different memory sizes, my bet is on ACPI. Soon the embargo will be lifted and we’ll find out!
A workaround in the beginning may be using WSL2.
Same, I have no interest in a Windows machine. First thing that I would want to do to an RTX is delete Windows and install Linux but I realize they’re just trying to capture the broader market.
Yeah, I have already seen comparisons and reviews. This thing is in price range of m5 max but unless you buy it for gaming (you should not, buy amd64 instead) m5 max destroys it in every single bench, as expected. It should be faster in raw GPU compute but even that is tanked by slow lpddr.
One thing I didn’t consider is that in the long run RTX might become much more supported and stable. And for most users who are tied with apps on windows you’d be able to get your portable assistant running on the same machine of your other apps, with no api nor internet access needed.
Macbook. Not even close to compare
Are you referring to my message? I don’t understand, in industrial setting I surely couldn’t use a Mac for everyday work related to the company.
Working in robotics and industrial settings I wouldn’t even consider using a Mac. Unless I can install Linux and Windows on that, but then why bother getting that hardware when you have much simpler alternatives?
That’s your constraint not mine. Rtx spark 128g is same ballpark price as m5 max, but Mac wipes the floor on any perf metric with it, especially Ai. And macos is basically 99% Linux. All my code developed on spark runs on MacOs without changes. Not the case in wsl2.
Excuse me?!? BSD is not Linux ;)
I don’t doubt that, and I don’t like working in Windows either (and I really don’t WANT to have anything to do with MacOS if I can avoid it, to be honest).
All I’m saying is that in most practical cases, at least here in Europe, neither MacOS nor Linux are viable alternatives in workplace environments. It’s not a judgement, it’s just what I see.
For example: I use Siemens TIA Portal, which is already a nightmare to install on Windows (I don’t even want to consider trying to make it work on Linux or MacOs), and it’s not that I have alternatives to it. Same goes for many automation and 3D modeling programs.
All I’m saying is: I initially evaluated RTX Spark only comparing it to DGX Spark and Apple alternatives on pure AI performance and scalability. But, in fact there’s a chance for RTX Spark to have a much bigger adoption at the same price level, just because it’s a Windows laptop. Bigger adoption means more support, more community contributions, and so on.
I’m not saying its the best platform, I’m pretty sure there are better price/performance alternatives. But it’s placed in a pretty unique position right now. It all really depends on which is the most easily usable alternative out of the box, not if I can squeeze 10 tps more at the same price.
Don’t care about church of linux, windows, macos. In all practical sense - linux stuff works on macos almost perfectly with very small exceptions. WSL2 despite being Ubuntu still rides on Windows kernel and throws feets at every opportunity, starting with UTF encoding ending with process supervisors and deeper stuff.
I really don’t understand the need of running local ai on the notebook in corporate settings. Never heard of that, would never approve my employees doing that. Try running a model taking 100-110GB out of 128B of ram and gunning GPU at 100% and see how workable this is. I use my macbook as a server but it sits on the crane with external cooling two ways, connected to hubs and multiple monitors. But nobody in corp does that, not have space/budget/approval.
Picture AI regulations coming up in Europe, users getting used to capable AI assistants, token spend frenzy having calmed down, corporate companies now steering towards local models due to privacy/data protection, and most public administrations, manufacturing and mainstream activities locked on Windows.
I’m not talking about how we’d use these machines, but how a less hardcore user would. Someone that has to do some plain old job and wants it to be simpler.
About the need of running local AI on the notebook in corporate settings, it’s all due to how things are progressing.
I’d never picture user running a 100-110 GB model on its 128GB machine, but I’d very well see a ~30-70B model MoE (Which, by the way, seems more balanced for the compute/bandwidth of Sparks), which would leave enough space for what that user would actually need to work on. IF it can be installed and run transparently to the user. And, maybe even fine-tuned on their specific niche?
And the reason why I see this coming is simple: AI assistants are EVERYWHERE now. I mentioned TIA Portal, right? For those who never had to do with it, it’s a huge clunky application that’s required to work with latest PLC models while keeping compatibility with 20+ years old devices, that requires a full day just to install its components in a fairly powerful machine just to run display an obsolete GUI that you’d expect on Windows 95.
Well, this nice piece of software now has its own AI assistant too!
The problem is: how many companies (corporate, public administration, hospitals, manufacturing companies, …) are ok giving away all its company data to AI giants? How many would trust the data not to go out anyway if those AI providers were to promise to keep it private?
We’ve been just through a “token spending frenzy” in corporate where everyone had to drain as much tokens as they could. Now, even taking away the spending, some companies realized that they were giving away their internal knowledge and secrets, and taking a step back toward other providers or local inference. But people is getting used to have capable assistants.
Once local inference is good enough (and we are about at that level now with models that fit 128GB, and smaller ones will be there soon), it’s quite clear that the most interesting option to have both capability and privacy will be having a model running on your machine.
Don’t get me wrong, I’m very well capable of setting up a cluster of machines for local inference and have them serve to anywhere I need. But us here don’t reflect the average AI user. And even then, even I could appreciate being able to use a fairly capable model on the same machine I need on my daily work with no connection required. Say, debugging a PLC program on a remote off-site location.
Obviously, for this to have any chance to pick up at scale we’d need a much more flawless user experience, including all the safety, privacy, stability and usability features that would have to come included.
I’m not saying this is the future, we are far from it. But with increasing demand I can picture this happen.
With Windows running I don’t see the ability to run a model that uses 100+GB when running on Linux.
I think the AI on the RTX Spark is more down the line of being a semi intelligent helper, with cloud AI (or DGX systems) still being used to do the heavy stuff.
Not 100+, but I have Qwen-3.8-Flash-Next running on RTX Spark using between 90-96 GB (which is the maximum) depending on KV cache. And that is a quite capable model. Not intending to use it like that, but it’s possible.
Yeah, I went from 1 to 2 DGXs when I exhausted the need for one. For me (and I am not discounting the platform) adding Windows on top, makes it just a play thing while I wait the results from the DGX cluster. So I will pass on RTX, understand the use case and capability, just not for me.