Support NeoGAF

[Wolzard]

Microsoft reports are exposing AI's real cost problem: Using the tech is more expensive than paying human employees | Fortune

Companies are racing to incentivize employees to use AI. But as some companies are finding, the more employees that use the technology, the heavier the bill.

fortune.com

In fact, this exposes the basics of any company. It's always more expensive to rely on third-party tools. That's why the headline is wrong. They stopped using another company's solution to use their own.

 

[twilo99]

Still far from Apple's monsters but good first try I guess

 

[Kenneth Haight]

This is for the bed rot gen, that's coming up.

 

[PaintTinJr]

I am curious on performance of these for local AI models and the price.

Unified memory will go a long way for the above potentially. However if it's more expensive than equivalent Mac MBPro or Studio, then meh.

How will that work with memory latency for CPU and GPU access?

I would have thought they would need a similar unified memory setup like the PS4/PS5 where they are setup as GDDR for high performance, higher latency with GPU access and then have a mode change when accessed with the CPU.

 

[delishcaek]

I wouldn't believe anything this clown says about gaming performance. Remember "4090 performance for $549".

Also... you could control your PC via voice commands before, this isn't new at all.

Use voice access to control your PC & author text with your voice | Microsoft Support

Learn how to use voice access for accessibility.

support.microsoft.com

 

[StereoVsn]

How will that work with memory latency for CPU and GPU access?

I would have thought they would need a similar unified memory setup like the PS4/PS5 where they are setup as GDDR for high performance, higher latency with GPU access and then have a mode change when accessed with the CPU.

They already have that with DGX Spark, Apple has had Unified Memory for ages, and AMD sort of has it for their AI 395 systems.

Wider bus and cache can allow decent enough bandwidth.

 

[Sethbacca]

It's interesting tech but i'm so burnt out on AI shit that I have zero interest.

 

[twilo99]

Essentially 2 generations behind the curve

 

[tronied]

Still far from Apple's monsters but good first try I guess

As far as I know the spark chip was never mean't to put itself as a leader in pure CPU compute. It's supposed to excel in AI workloads where it can offload it from the CPU onto the tensor cores. As such, without providing benchmarks using local LLM's / models, it's a bit of a mute point. I guess it's impressive that a newcomer coming to the stage can put up those numbers. Will be interesting to see how follow-up iterations progress. Not really that interested myself as running local LLM's is not really up high on things I'm looking for currently.

 

[Brucey]

Essentially 2 generations behind the curve

Like the t239 for the Switch 2.....nvidia found another turd to polish and send out this time for windows market.

 

[xrnzaaas]

He's going for that 'most hated tech company CEO' trophy.

 

[Loope]

O god when does it stop



This is literally how a PC has been operating since its inception: you give a command and the PC does the thing.

 

[Reizo Ryuu]

"beep glorp brrrrr *fart*," said Jensen Huang,

Yep

 

[twilo99]

As far as I know the spark chip was never mean't to put itself as a leader in pure CPU compute. It's supposed to excel in AI workloads where it can offload it from the CPU onto the tensor cores. As such, without providing benchmarks using local LLM's / models, it's a bit of a mute point. I guess it's impressive that a newcomer coming to the stage can put up those numbers. Will be interesting to see how follow-up iterations progress. Not really that interested myself as running local LLM's is not really up high on things I'm looking for currently.

True, I should've said "in certain aspects" but still shows how good the apple chips are overall.

It's a solid push as far as ARM and windows

They will even have Jensen on stage at build this year I think that's a first

 

[Tommi84]

How many body organs would I have to sell to be able to afford this?

 

[Zacfoldor]

Think of this in videogame terms.

Most of you, oddly, are luddites. Probably because you have families(just a guess). Again, think of this like it was Marathon.

People, not luddites, who use AI will have a tremendous advantage in everything they can do or accomplish vs people who refuse to adopt it. Over time this gap will grow exponentially. AI is basically a superpower, the bat computer. If nothing else the ability of it to automate processes puts users far about the older style of human who does everything manually one problem at a time.

So you people trying to slow this down because you have families, you are fucking up.

Instead, make sure your families are adopting this now. Otherwise they won't compete in the market of the future and the skill/knowledge gap between then and
AI adopters will be tremendous. If you are treating this like NFTs and putting your head in the sand you are missing the boat. Believe me. Get in on this now. Early adopters will be rewarded, and it's just a tool. It is not at all as evil as you imagine. Avoiding it now is like refusing to use the Internet during the .com boom because it took away money from local book and pet stores.

 

[mrBarrelNut]

The price point will be hilarious

On par with high end Macbooks I'm sure.

 

[Magic Carpet]

On par with high end Macbooks I'm sure.

Macbook Neo using that iphone a18pro which google ai is calling a 'baby M4'.
It's only $600

 

[Exentryk]

Definitely interested in this type of unified memory setup on Windows. While Macbooks are great for AI, they do run into various challenges in some areas like Stable Diffusion. Being able to run the Nvidia stack with large unified memory is cool.

Will probably consider these in a few more generations once the bandwidth is faster.

 

[StereoVsn]

Essentially 2 generations behind the curve

On one hand "Oof" as M5 is about 30-40% ahead on CPU performance. On the other hand most of AI efficacy is on GPU side so that's going to be most important feature set to watch.

Apple has been killing it recently and this competition is going to be interesting to watch.

 

[twilo99]

On par with high end Macbooks I'm sure.

Even if they were going after that for running local LLMs the high end Mac still has faster memory so charging the same wouldn't be wise ..

 

[Brucey]

Think of this in videogame terms.

Most of you, oddly, are luddites. Probably because you have families(just a guess). Again, think of this like it was Marathon.

People, not luddites, who use AI will have a tremendous advantage in everything they can do or accomplish vs people who refuse to adopt it. Over time this gap will grow exponentially. AI is basically a superpower, the bat computer. If nothing else the ability of it to automate processes puts users far about the older style of human who does everything manually one problem at a time.

So you people trying to slow this down because you have families, you are fucking up.

Instead, make sure your families are adopting this now. Otherwise they won't compete in the market of the future and the skill/knowledge gap between then and
AI adopters will be tremendous. If you are treating this like NFTs and putting your head in the sand you are missing the boat. Believe me. Get in on this now. Early adopters will be rewarded, and it's just a tool. It is not at all as evil as you imagine. Avoiding it now is like refusing to use the Internet during the .com boom because it took away money from local book and pet stores.

Your constant shilling for AI is impressive.

 

[StereoVsn]

Your constant shilling for AI is impressive.

Yeah, I have yet to see anyone outside Tech Bros in SF to be so into AI spiel.

And I say this as someone who is running one of main AI initiatives for my company and a heavy user of AI at home for personal and gig projects.

 

[twilo99]

Did they mention anything about "Xbox mode" support with this thing?

It's probably a good opportunity to add some back compatibility layers if they wanted to push some unique "gaming" features on it

 

[Brucey]

Did they mention anything about "Xbox mode" support with this thing?

It's probably a good opportunity to add some back compatibility layers if they wanted to push some unique "gaming" features on it

Sounds like there may be a handful of "native games/re-releases" written for the ARM cores.

 

[StereoVsn]

Did they mention anything about "Xbox mode" support with this thing?

It's probably a good opportunity to add some back compatibility layers if they wanted to push some unique "gaming" features on it

Nah, but XBOX is all in!

 

[ResilientBanana]

An AI chip? whaaat no waaay.

Essentially 2 generations behind the curve

To be fair, it's version 1. nVidia doesn't like being at the bottom of the totem pole. Secondly, these scores are from 2025, optimization of drivers and firmware might provide significant improvements.

 

[Hohenheim]

"This is the new PC. The personal AI computer."

No thanks, Jensen. I already deactivate every trace of copilot and AI-bloat in my pc's, and this is not for me.
I'll rather pay overprice for tradiotional "old-skool" cpu's and gpu's and keep my pc's free of all this shit they keep throwing at us.

 

[StereoVsn]

An AI chip? whaaat no waaay.


To be fair, it's version 1. nVidia doesn't like being at the bottom of the totem pole. Secondly, these scores are from 2025, optimization of drivers and firmware might provide significant improvements.

Keep in mind that both CPU and GPU on M5 is quite a bit faster.

I think the big advantage is CUDA but disadvantages are also many. Either way this will be interesting to see when it releases.

 

[memoryman3]

Introducing a powerful new chapter for Windows PCs, accelerated by NVIDIA RTX Spark

Today at NVIDIA GTC, Microsoft and NVIDIA announced the world’s most powerful and efficient thin-and-light Windows PCs ever. Accelerated by NVIDIA RTX Spark

blogs.windows.com

About gaming:

Prism emulation enhancements

Prism, our emulator for running 32-bit and 64-bit x86 apps on Windows on Arm, will also be present and optimized for RTX Spark powered PCs. Prism ensures apps run well on these devices even if those apps haven't been built for the Arm architecture. We have continued to enhance the Prism emulator with additional performance and compatibility features, building on the Prism optimizations delivered last year that added support for the AVX/AVX2 instruction set extensions. Prism has been tuned for the microarchitecture of RTX Spark and when combined with the raw power of the silicon, unlocks great performance for developers, creators and gaming workloads running under emulation.

On Game Developers Support

Game developers have also laid a strong foundation for RTX Spark's arrival. Today, native anti-cheat solutions from partners like Epic's Easy Anti-Cheat and BattlEye, expanded Prism emulator compatibility, and XBOX PC app support means players will have access to a deep catalog of Windows PC games. RTX Spark will bring even higher levels of gaming performance to AAA titles on Arm. Riot Games, one of the world's leading game developers and publishers, has announced that League of Legends and VALORANT are coming to the platform. PUBG: Battlegrounds, the iconic battle royale title from KRAFTON, will also be joining the expansive catalog of compatible titles including Pragmata, Alan Wake 2*,* Naraka: Bladepoint, War Thunder and more.

---

This first generation should focus on creators and developers, but the idea seems to be to create a new market within the ARM architecture, especially since Nvidia has shown that it already has two more generations planned.

This is an Xbox.

 

[64gigabyteram]

Pushing PC laptops into ARM should've been a slam dunk with Nvidia help but because of this massive AI side quest it's gonna be underwhelming and overpriced to anybody who doesn't GAF about "Claude"

 

[diffusionx]

An AI chip? whaaat no waaay.


To be fair, it's version 1. nVidia doesn't like being at the bottom of the totem pole. Secondly, these scores are from 2025, optimization of drivers and firmware might provide significant improvements.

The issue is that they are charging near Apple prices for an AI laptop that isn't as good as the Apple. A couple hundred bucks saving isn't going to make a difference to anyone who wants to buy a laptop for AI.

Like ok it's a good start but it's DOA.

 

[mrBarrelNut]

Essentially 2 generations behind the curve

CPU performance is not the focus on these devices, the GPU + CUDA is.

 

[HerjansEagleFeeder]

This global GPU de-facto monopoly has become very enjoyable.

 

[Topher]

This is an Xbox.

 

[PaintTinJr]

They already have that with DGX Spark, Apple has had Unified Memory for ages, and AMD sort of has it for their AI 395 systems.

Wider bus and cache can allow decent enough bandwidth.

A quick chat with AI suggests the N1X CPU side bandwidth will be high-end mobile - consistent with the DGX - because of the LP DDR choice rather than desktop level CPU memory bandwidth that PS5/XsX approach, and the 300GBs figure is probably true for the GPU side. The AI doesn't believe the unified memory will be able to dereference unified memory between CPU and GPU directly like the PS4/PS5 can, and will likely use a hardware link + software stack but still need to go through the OS mechanisms for indirect unified memory essentially leaving a little performance left on the shelf with regards to memory.

 

[StereoVsn]

A quick chat with AI suggests the N1X CPU side bandwidth will be high-end mobile - consistent with the DGX - because of the LP DDR choice rather than desktop level CPU memory bandwidth that PS5/XsX approach, and the 300GBs figure is probably true for the GPU side. The AI doesn't believe the unified memory will be able to dereference unified memory between CPU and GPU directly like the PS4/PS5 can, and will likely use a hardware link + software stack but still need to go through the OS mechanisms for indirect unified memory essentially leaving a little performance left on the shelf with regards to memory.

That's very likely but in general should be ok-ish. Apple will have significantly higher bandwidth on M5 Max and upcoming M5 Ultra.

 

[CobraAB]

I see nothing below $2999 here and I am probably way to optimistic...

 

[IntentionalPun]

In fact, this exposes the basics of any company. It's always more expensive to rely on third-party tools. That's why the headline is wrong. They stopped using another company's solution to use their own.

Microsoft's own solution relies heavily on Anthropic's models. They want people to be testing their tooling (GitHub Co-Pilot w/ VS or VS Code) but they are still paying Anthropic loads of money.

And anyone using GitHub Co-Pilot is making Anthropic money. Most complex tasks are routed through Anthropic's models, and beyond that you can select them individually and most do for any software dev.

 

[Buggy Loop]

As far as I know the spark chip was never mean't to put itself as a leader in pure CPU compute. It's supposed to excel in AI workloads where it can offload it from the CPU onto the tensor cores. As such, without providing benchmarks using local LLM's / models, it's a bit of a mute point. I guess it's impressive that a newcomer coming to the stage can put up those numbers. Will be interesting to see how follow-up iterations progress. Not really that interested myself as running local LLM's is not really up high on things I'm looking for currently.

Yea these reactions make no sense

geekbench score, really?

That's not what this APU set out to do. When you'll benchmark for LLM performances it will go way differently. Gaming performances too.

128GB RAM will kill off the price though. Only really dedicated local LLM geeks will think of this product.

 

[IntentionalPun]

Yea these reactions make no sense

geekbench score, really?

That's not what this APU set out to do. When you'll benchmark for LLM performances it will go way differently. Gaming performances too.

128GB RAM will kill off the price though. Only really dedicated local LLM geeks will think of this product.

This APU is certainly going to be used for more than running local LLMs so I'm not sure why a GeekBench score wouldn't be relevant, especially one related to code compilation.

Everyone wants to compete with Apple in this space and the M-series are used by software developers heavily as well as other smaller markets like video/photo editing professionals.

 

[Topher]

That's very likely but in general should be ok-ish. Apple will have significantly higher bandwidth on M5 Max and upcoming M5 Ultra.

M5 Max bandwidth is between 460 GB/s and 614 GB/s. RTX Spark is 300 GB/s.

Will be interesting to see how much that translates into real world performance for local AI. Potential bottleneck for Nvidia

I wonder if these are locked down and you can only install Windows

 

[Buggy Loop]

This APU is certainly going to be used for more than running local LLMs so I'm not sure why a GeekBench score wouldn't be relevant, especially one related to code compilation.

Dude, if you're gonna drop that kind of money with laptops that have 128 GB RAM in these market conditions, NPU focused architecture, it's for LLMs and not for anything else.
It's a DGX spark packaged differently. Only LLM geeks will even look at the store page for that and it depends on what they want to do.

Everyone wants to compete with Apple in this space and the M-series are used by software developers heavily as well as other smaller markets like video/photo editing professionals.

Oh I'm sure there's a lot of artists wannabes that bought an M5max because.... "apple". Nvidia is not competing here for that. This is purely agentic focused laptop.
Peoples buying an M5 max with $1500 adder for 128 GB and using it for something else than LLMs, well they have more money than common sense. The average apple ecosystem

 

[StereoVsn]

M5 Max bandwidth is between 460 GB/s and 614 GB/s. RTX Spark is 300 GB/s.

Will be interesting to see how much that translates into real world performance for local AI. Potential bottleneck for Nvidia

I wonder if these are locked down and you can only install Windows

I hope it's not just Windows as Linux would be a must for a lot of AI work.

Also, M5 Ultra is going to be super interesting, might break 1TB on bandwidth (price will be insane).

 

[StereoVsn]

Dude, if you're gonna drop that kind of money with laptops that have 128 GB RAM in these market conditions, NPU focused architecture, it's for LLMs and not for anything else.
It's a DGX spark packaged differently. Only LLM geeks will even look at the store page for that and it depends on what they want to do.



Oh I'm sure there's a lot of artists wannabes that bought an M5max because.... "apple". Nvidia is not competing here for that. This is purely agentic focused laptop.
Peoples buying an M5 max with $1500 adder for 128 GB and using it for something else than LLMs, well they have more money than common sense. The average apple ecosystem

M5 Max boxes are also quite good for video editing or 3D rendering. But yeah, now days it's probably mostly AI work.

 

[lastmessiah]

I hope it's not just Windows as Linux would be a must for a lot of AI work.

Also, M5 Ultra is going to be super interesting, might break 1TB on bandwidth (price will be insane).

Windows has a fully featured Linux subsystem: https://en.wikipedia.org/wiki/Windows_Subsystem_for_Linux

 

[Hudo]

More like "uninvent".

 

[StereoVsn]

Windows has a fully featured Linux subsystem: https://en.wikipedia.org/wiki/Windows_Subsystem_for_Linux

I know and I have been using it for years. However it's not quite the same.

 

[lastmessiah]

I know and I have been using it for years. However it's not quite the same.

for dev work it is really all you need in my experience.

 

[Dorfdad]

The price point will be hilarious

599.99 bbbeeeeeeellllliiiiivvvvvvvvveeeeeeee

On a serious note I wonder if any of this makes its way to GeForce now!?