Dell confirms XPS laptop with NVIDIA N1X at Computex ( basically a DGX Spark GB10 for consumers with Windows )

[-]

Everyone_dreams@reddit

My slightly older XPS 13 is a freaking furnace on just normal task like video play back and simple games. I can only imagine this thing melting trying to run anything this sub would want to try on it.

[-]

dylanger_@reddit

Everyone as a whole needs to move away from Windows.

First class Linux support would be much better for this.

[-]

dryadofelysium@reddit

Why do you think these would have any issues running Linux? Should be fine.

[-]

FullstackSensei@reddit

Laptops are much harder to add support for if the maker doesn't do the work.

[-]

dryadofelysium@reddit

Dell usually has good support at least for Ubuntu with their own repo, and even my Microsoft Surface devices run fine nowadays

[-]

FullstackSensei@reddit

How long did it take your surface laptop to have support? and what is the definition of fine? Does everything work the same? Do you get the same or better battery life on Linux? No issues with suspend or power management?

The thing you and many don't seem to grasp is: ARM SoCs are very different from x86. In ARM land, you're completely dependent on the SoC maker releasing and updating device trees and drivers. Nvidia's history here, with Jetson boards, is far from stellar.

You're also ignoring the whole angle Nvidia themselves are pitching here: Windows on ARM, hence the copy-paste tweets from Nvidia and Microsoft about Computex.

The current crop of Snapdragon X Elite ARM laptops have seen lackluster support on Linux. Heck, my old fart HP Elite Folio with an 8CX from 2021 still doesn't have Linux support.

[-]

dryadofelysium@reddit

> and what is the definition of fine?

Fine means that on my Surface Laptop Studio everything works OOTB with vanilla Ubuntu except for the touchscreen, which requires installing the surface-linux kernel.

> The thing you and many don't seem to grasp is: ARM SoCs are very different from x86

I am literally earning my money building software on and for both x86 and ARM platforms. Yeah, I know how ARM works.

> You're also ignoring the whole angle Nvidia themselves are pitching here: Windows on ARM

My initial post was literally a reply to the comment "First class Linux support would be much better for this.". Are you okay? I hope you are okay.

> The current crop of Snapdragon X Elite ARM laptops have seen lackluster support on Linux.

I am well aware, which everyone building Android ROMs for the past 15 years could have seen from a mile away and which is why there is some interest in seeing other SoC vendors, even if there will be closed source blobs involved.

[-]

LetsGoBrandon4256@reddit

How's the battery life when compared to Windows?

[-]

dryadofelysium@reddit

No difference. That being said, even though I am a Microsoft Surface fanboy, I will say the battery has never been great on hardware from Covid times and before.

[-]

SkyFeistyLlama8@reddit

What's with the Linux evangelists on here to begin with? They're bad enough in the ThinkPad sub, ffs.

For it to be a consumer or enterprise laptop, it has to run Windows or Windows on ARM in this case. No big deal, Qualcomm has been paving the way since 2018 and I can say Windows ARM is pretty much mature at this stage.

Linux means some flavor of Linux being maintained by Nvidia. As you said, Nvidia has a terrible track record of supporting SBCs and non-data center hardware on Linux.

All that being said, I'm skeptical about what Nvidia and Microsoft can achieve together with an N1X Windows on ARM laptop. What do you need that much RAM and GPU power for in that form factor (other than machine learning)?

[-]

ahjorth@reddit

Running local models requires robust and well-supported toolchains, efficient implementations, and most people here have a strong preference for privacy.

Windows doesn't have the same suite of toolchains, it runs slower, and is slowly but unambiguously eroding opt-out options for telemetry.

I think you are having a very strong reaction to a completely reasonable stance.

[-]

FullstackSensei@reddit

I have a snapdragon powered WoA laptop and it's far from mature. If all you care about is mail, browser and office, you're mostly right. But if you or your business have or built some custom application, there's a good chance said application doesn't work or doesn't work well on WoA, not to mention the lack of stability in x86 emulation.

As for why one would need that much GPU power, it's the same market usually addressed by mobile workstations: engineering, image and media editing, etc.

The N1x is Nvidia's answer to Strix Halo. The question is, given how expensive everything related to computers is and the generally shitty state of Windows 11, how many businesses will take a risk on such an unproven, yet expensive, platform?

[-]

NoleMercy05@reddit

So fix Linux

[-]

bnolsen@reddit

The next iteration of this type of system released by Nvidia will likely see it abandoned support wise. On the other hand strix halo support will continue improving for years and will the system usable for many years.

[-]

relmny@reddit

First thing I thought when I read the tittle was "ah, so not aimed for local AI..." because there Linux have way better support for AI than windows.

[-]

Technical-Earth-3254@reddit

256GB unified RAM Systems would be great

[-]

absurd-dream-studio@reddit

You will only get 256 gb bandwidth

[-]

stonktraders@reddit

Probably around 270GB/s with 256bit LPDDR5X

[-]

soyalemujica@reddit

Sad, so stuck to about 9t/s with Qwen 3.6 27B dense for example

[-]

Eastern_Bet678@reddit

Can it ... not run Windows?

[-]

SeyAssociation38@reddit

No, it's aimed at the consumer market.

[-]

relmny@reddit

And clearly not at the AI consumer market...

[-]

SeyAssociation38@reddit

It aimed at the average high end consumer who would run ollama

[-]

m7l5@reddit

Dell officially supports Ubuntu (or vice versa but you get proper drivers for their laptop which is good). So I guess yes.

[-]

Maleficent-Ad5999@reddit

There was a rumour that n1x was delayed because they couldn’t get shit done working with windows.. if they had focused on making it for Linux , we would have had a 2nd gen by now

[-]

CatalyticDragon@reddit

You can run linux but bugs and issues might be even worse and Microsoft's Prism x86 emulator could be better than alternatives for linux.

- https://github.com/NVIDIA/dgx-spark-playbooks/issues

- https://docs.nvidia.com/dgx/dgx-spark/known-issues.html

- https://docs.nvidia.com/dgx/dgx-os-7-user-guide/known_issues.html

Obviously it's a wait and see situation.

[-]

SeyAssociation38@reddit

Honestly they seem pretty tame, few and far between and "known" unlike Windows. Dgx aka Linux is already so much better than windows. It should come with it pre installed

[-]

Original_Finding2212@reddit

SparkOS?

[-]

nekmatu@reddit

Just a warning. Dell XPS have been absolute dogshit lately with throttling issues due to heat. We bought a bunch for my team last year. Top of the line, fully specced and they run slower than 5 year old computers. Constantly throttling, freezing for no reason.

I wouldn’t by another XPS until they’ve decided to actually build something decent.

[-]

bnolsen@reddit

Sounds like the newer macbooks.

[-]

nekmatu@reddit

Interesting. Didn’t know there were issues with those. We got the m5s now and no complaints so far.

[-]

ImaginationKind9220@reddit

If it's sold at the same price as DGX Spark, then it will be a flop. You can buy a 5090 laptop at a cheaper price and it's faster.

[-]

sn2006gy@reddit

I can't imagine the market is DGX Spark pricing, this has to be something competing against the 600 dollar apple. There is no way MIcrosoft would be interested in killing its surface premium brand with something more expensive.

[-]

JaredsBored@reddit

That’s not at all what this is. This'll be a 128GB of ram MacBook pro competitor. Microsoft doesn't have anything in this tier. This is high end (for laptop) compute. And besides, Microsoft has the Surface brand to push laptop makers to make higher quality hardware, big they still make their money on software.

Microsoft has wanted windows on arm to be a "thing" for a decade plus. Qualcomm is too busy shitting the bed drivers-wise to make any headway, and Nvidia is learning from that mistakes. They've been sitting on this product launch for months while they get the drivers perfected* (we'll see if they're actually perfect, highly doubt it).

[-]

sn2006gy@reddit

Microsoft has had ARM for a while, the Original Surface RT was Nvidia. If this is a 3.5k laptop, it's dead on arrival

[-]

thrownawaymane@reddit

Yes, but they only just committed fully to ARM as a first class citizen.

Up until last year you couldn't even download an ISO of it and the number of third party hoops you needed to jump through to make one was impressive

[-]

sn2006gy@reddit

They've had a Qualcomm developers kit for years now and a surface pro x ARM variant for years (and multiple releases/versions)

[-]

thrownawaymane@reddit

Yes, and I support several of those devices at work.

Be happy you didn't use that dev kit at the beginning...

[-]

sn2006gy@reddit

I did use the devkit, it was a cheap way to get ARM and I never had driver problems, SoC systems tend to actually be pretty stable because they don't need complex driver subsystems since the SoC is fully integrated. The Surface Pro X has been a champ and not once did i ever concern myself with drivers.

[-]

fallingdowndizzyvr@reddit (OP)

Yes, but they only just committed fully to ARM as a first class citizen.

The committed to that with Qualcomm.

[-]

Apprehensive-View583@reddit

128gb max has twice plus of memory bandwidth, I mean if it will have similar bandwidth it might be able to sell, if it’s gb10 bandwidth, I can’t think of any reason not to buy a Mac..

[-]

sn2006gy@reddit

Correct, this laptop doesn't lign up with premium macbook pro specs unless they're doubling the ram bandwidth in which case, RIP Spark.

[-]

JaredsBored@reddit

We need to get away from assuming that memory bandwidth is the whole story. It's *very* important, don't get me wrong, but the spark has a significant prefill advantage. From the numbers I'm seeing online, it's significantly faster in prefill using vLLM on spark vs MLX on MacBook.

MacBook does have a token gen advantage advantage, but it's not double. The available software stack to each platform matters as much as the hardware.

Qwen 122b 4bit, concurrency 1, 0 depth prompt processing (imperfect because the spark did 2048 but the m5 did 4096, but close enough):
Spark: 2124tps
M5 Max 40c: 1517tps

Qwen 122b 4bit, concurrency 1, 4096 depth token gen:
Spark: 43tps
M5 Max 40c: 62tps

Qwen 27b 4bit, concurrency 1, 0 depth prompt processing (imperfect because the spark did 2048 but the m5 did 4096, but close enough):
Spark: 1792tps
M5 Max 40c: 850tps

Qwen 27b 4bit, concurrency 1, 4096 depth token gen:
Spark: 22tps
M5 Max 40c: 32tps

Sources:

https://omlx.ai/benchmarks/q3c89xz3
https://omlx.ai/benchmarks/9yrp63om
https://spark-arena.com/benchmark/da895d78-e532-4fd8-a31b-24a3392313f0
https://spark-arena.com/benchmark/440ad5de-83c3-4ac4-b07b-dbdd395fb7dc

[-]

sn2006gy@reddit

You write as if i don't already own and use both. I've already said the sweet spot for this chip would be MOE - certainly not 27b and if you wanted to be real about the 27b the macbook would smoke the spark on real use because the spark collapses to 4-5tps with any context at all but this is all beside the point. I own Spark and a MacBook - Microsoft can't put this chip up against the macbook pro unless they do address the memory issue - they just can't.

MediaTek has never been known for pushing performance hard and while the GB10 is nice, it's held back by some massive pifalls still not addressed in software (no native NVFP4 kernels) and i don't see that being fixed on Windows to push a 3.5k price on these. I just don't.

[-]

Serprotease@reddit

The spark gets notably warm and pulls 100w+ from the wall in these conditions. It’s quite likely that the power available to the gpu will be lower with additional thermal constraints on a laptop chassis. As much as I like the spark, the reason to get one (small/cluster potential with tensor parallelism, good processing power, cuda) may not hold here.

And, you may be forced to deal with windows if Nvidia doesn’t provide proper Linux drivers support.

[-]

SkyFeistyLlama8@reddit

How is Qualcomm shitting the bed with drivers? Adreno runs fine on WoA. Gaming support is improved on X2 compared to X1.

[-]

JaredsBored@reddit

Here's the whole article rundown on the driver mess on Linux: https://www.phoronix.com/review/snapdragon-x-elite-linux-eoy2025

[-]

Deep90@reddit

Could be a "M" macbook competitor to start, no?

XPS laptops are usually flagships.

[-]

fallingdowndizzyvr@reddit (OP)

But Dude it's a Dell! I've bought Dell laptops over the years. Sooner or later, Dell laptops are sold cheap. Like super cheap. They are heavily discounted S&D. Even if they don't have noticeable S&D. I've gotten S&D Dells that have been pristine. So even if it's initial price is high, sooner or later it's going to be bargain basement with hopefully a pretty substantial Amex statement credit you can apply. Probably sooner rather than later. Especially if it's a flop.

[-]

SkyFeistyLlama8@reddit

It won't be sold cheap. I don't see any of the DGX OEM units being sold for a lot less than Nvidia's prices. This device isn't some cheap XPS.

This will be a unicorn: a portable AI workstation with CUDA and lots of RAM, something that even Apple doesn't have.

[-]

fallingdowndizzyvr@reddit (OP)

This will be a unicorn: a portable AI workstation with CUDA and lots of RAM, something that even Apple doesn't have.

But that's not it's intent. It's intent is to bring Blackwell to the masses. Or should I say other masses. Since the 5000 series GPUs already do bring Blackwell to the masses.

This is just a XPS laptop like plenty of others. If it was intended to be a unicorn they would have new branding.

HP also has a premium Strix Halo laptop. It was also a $4000-$5000 laptop. A month or so ago, it was sold for $3000 for the high end config with the OLED display and 128GB of RAM. IMO, this XPS competes with that HP.

[-]

SkyFeistyLlama8@reddit

AMD and Intel laptops don't command a premium for long. They're only expensive at launch and prices slowly come down. A Strix Halo laptop unfortunately doesn't know what it wants to be: a thin and light workstation, a premium gaming laptop or an AI toy with nerfed software frameworks. AMD doesn't know what to do with it and OEMs like Dell also don't know what to do with it.

Compare that to a MacBook Pro that has premium pricing that rarely ever moves downwards. This N1X laptop is competing against the MacBook Pro with the added bonus of having CUDA and 128 GB RAM onboard. That for me blows the M5 MBP Max out of the water already.

[-]

Serprotease@reddit

I would like to think that it will better than a M5max, but it’s certainly losing on the cpu side (M5max > gb10) and being in a laptop format, thermal and power draw will be an issue.

Probably interesting for diffusion models and nice if it can run Linux. But for Llm? Unlikely.

[-]

SkyFeistyLlama8@reddit

ARM Linux on a laptop? Yeah wait for 2030. The DGX already has to run an Nvidia-specific Linux build. Add laptop bits and parts and see compatibility issues go right through the roof.

Thermals and power will be an issue so it could be limited to max 150W or 200W. Then again, you have a decent amount of prefill speed coupled with full CUDA compatibility to work with models directly instead of hoping for second-tier frameworks like MLX or ROCM.

[-]

Serprotease@reddit

It seems that both Lenovo and dell are looking to integrate it into either thin laptops (yoga/xps) or gaming one (legion) so we could see a wild range of performance based on the power allocation. Probably not more than 60/70w on the thin ones (same as some of the 4070 laptops).

[-]

fallingdowndizzyvr@reddit (OP)

A Strix Halo laptop unfortunately doesn't know what it wants to be: a thin and light workstation, a premium gaming laptop or an AI toy with nerfed software frameworks.

Ah... an HP Zbook Ultra knows exactly what it is. A premium workstation laptop. The same as a Dell XPS. The Macbook Pro is the competitor to the HP Zbook Ultra. That's why the two are often compared.

[-]

Alive_Wedding@reddit

It will probably have substantially less RAM

[-]

SangerGRBY@reddit

M5 macbook pro would be better no ? GB10 is slower than M5 ?

[-]

NoleMercy05@reddit

No

[-]

SangerGRBY@reddit

Isit not ? I recall watching some video that M5s are faster than GB10.

[-]

MajorZesty@reddit

From the reviews I've seen on Dell XPS models over the years, I'd expect it to be very throttled. It's a weird choice, since Dell sends to limit the TDP in XPS a lot lower than other models.

The likely top configuration based on an Nvidia N1X 675 is listed at the equivalent of €4,049, with 64GB RAM also included. A slightly cheaper version with an Nvidia N1X 650 APU and 32GB RAM comes in at €3,199.

https://www.notebookcheck.net/New-PC-era-Lenovo-laptop-leak-unveils-first-Nvidia-N1X-notebooks-and-high-prices.1311489.0.html

[-]

nekmatu@reddit

This is correct. We bought 20. All dog shit. Throttle and freeze all the time. Engineers fucking hate them and went back to their old computers.

[-]

Ok_Needleworker_6431@reddit

The missing number is memory bandwidth. That decides tokens/sec, not the 6144 CUDA cores. Reserving judgment until someone posts actual llama.cpp throughput - still a cool start to something meaningful in the consumer space tbh.

[-]

fallingdowndizzyvr@reddit (OP)

That decides tokens/sec, not the 6144 CUDA cores.

Which tokens/sec. PP or TG. Since it's those cores, that compute that decides it for PP.

Reserving judgment until someone posts actual llama.cpp throughput

There are already plenty of posts about GB10 performance. Yes, it's light on TG if you are running tiny models. But it's pretty darn good in PP even comparing it to a tiny VRAM GPU.

[-]

j_osb@reddit

Just to chime in, IIRC it's roughly 5070 class in FP16 tensor with FP16 accumulate.

So slightly worse than 5070, and worse than 5070ti, 9070/xt in PP.

Roughly 15% slower than a 3090 in PP at FP16.

[-]

n4pst3r3r@reddit

GB10 has about 300GB/s

Would be solid for running small-ish MoEs. Qwen 122B-A10B at Q4 could theoretically yield 60t/s.

https://www.techpowerup.com/340385/nvidia-dissects-gb10-superchip-soc-with-20-cpu-cores-and-6-144-cuda-gpu-cores

[-]

Ok_Needleworker_6431@reddit

Didn't know this, thank you for sharing.

[-]

Zeeplankton@reddit

I think given how the spark shit the bed in memory bandwidth I don't see this being better. I think this will just be a consumer chip to take on apple.

[-]