MrRibeye
Member
Imagine if Apple was like "The new MacBook with ™, ™, ®, ™ and ®."
-
Hey Guest. Check out your NeoGAF Wrapped 2025 results here!
NVIDIA and Microsoft Reinvent Windows PCs (laptops) for the Age of Personal AI
- Thread starter cormack12
- Start date
Imagine if Apple was like "The new MacBook with ™, ™, ®, ™ and ®."
StereoVsn
Gold Member
Yeah, but if you are going to go into heavy LLM work I wouldn't think you would want to mainly utilize WSL.
Buggy Loop
Gold Member
Bandwidth ain't everything when even the M3 ultra had ~819 GB/s 512 GB
It's good for generation time. Apple M's are weak on compute side.
- Prefill → high compute device.
- Decode → high memory-bandwidth device.
Good benchmarking and explanation here.
Combining NVIDIA DGX Spark + Apple Mac Studio for 4x Faster LLM Inference with EXO 1.0
Disaggregating Prefill and Decode: Faster First Tokens, Faster Streams
blog.exolabs.net
So they are apples and oranges.
NVIDIA DGX Spark vs Mac Studio M5 Max for Local AI 2026 | Presenc AI
Head-to-head 2026 comparison of NVIDIA DGX Spark and Mac Studio M5 Max 128GB for local LLM inference, fine-tuning, and on-device RAG: throughput, memory...
A complex agentic coding tasks where the agent has a big system prompt, reads in tool inputs and files, will perform better on Spark. If you want to actually train and fine-tune models (not just run them), the spark wins on compute alone.
These products are so niche and financially doesn't even make sense when you can probably buy a decade of cloud AI that will continue to advance at a rapid pace, over buying any of these solutions for local LLM, imo. Unless you're generating things that clouds would stop you doing... those peoples.
The DGX spark actually made sense for network engineers to dev time on it locally and knew it was the same output for the datacenters. Not sure who will spend that kind of cash on that laptop.
MCH2024
Member
Guys remember N1X is basically the same power as Panther Lake but 65% faster or so.
18 months late and still it leaves no survivors. This is because Intel and AMD are only measured against each other. But they have fallen significantly behind ARM Trio due to incompetence (Intel) or lack of caring about Client (AMD).
On the graphics side, RTX has always had 2-6 years of leadership over AMD.
The end result is a product delayed 18 months not only being competitive but dominating.
Last edited:
Bandwidth ain't everything when even the M3 ultra had ~819 GB/s 512 GB
It's good for generation time. Apple M's are weak on compute side.
DGX wins in raw compute and FP4 allowing large models. It's prefill performances likely not beaten until M6, if even that.
- Prefill → high compute device.
- Decode → high memory-bandwidth device.
Good benchmarking and explanation here.
Combining NVIDIA DGX Spark + Apple Mac Studio for 4x Faster LLM Inference with EXO 1.0
Disaggregating Prefill and Decode: Faster First Tokens, Faster Streamsblog.exolabs.net
So they are apples and oranges.
NVIDIA DGX Spark vs Mac Studio M5 Max for Local AI 2026 | Presenc AI
Head-to-head 2026 comparison of NVIDIA DGX Spark and Mac Studio M5 Max 128GB for local LLM inference, fine-tuning, and on-device RAG: throughput, memory...presenc.ai
A complex agentic coding tasks where the agent has a big system prompt, reads in tool inputs and files, will perform better on Spark. If you want to actually train and fine-tune models (not just run them), the spark wins on compute alone.
These products are so niche and financially doesn't even make sense when you can probably buy a decade of cloud AI that will continue to advance at a rapid pace, over buying any of these solutions for local LLM, imo. Unless you're generating things that clouds would stop you doing... those peoples.
The DGX spark actually made sense for network engineers to dev time on it locally and knew it was the same output for the datacenters. Not sure who will spend that kind of cash on that laptop.
So the dgx spark is a screamer fan noise wise?
MCH2024
Member
Bandwidth ain't everything when even the M3 ultra had ~819 GB/s 512 GB
It's good for generation time. Apple M's are weak on compute side.
DGX wins in raw compute and FP4 allowing large models. It's prefill performances likely not beaten until M6, if even that.
- Prefill → high compute device.
- Decode → high memory-bandwidth device.
Good benchmarking and explanation here.
Combining NVIDIA DGX Spark + Apple Mac Studio for 4x Faster LLM Inference with EXO 1.0
Disaggregating Prefill and Decode: Faster First Tokens, Faster Streamsblog.exolabs.net
So they are apples and oranges.
NVIDIA DGX Spark vs Mac Studio M5 Max for Local AI 2026 | Presenc AI
Head-to-head 2026 comparison of NVIDIA DGX Spark and Mac Studio M5 Max 128GB for local LLM inference, fine-tuning, and on-device RAG: throughput, memory...
A complex agentic coding tasks where the agent has a big system prompt, reads in tool inputs and files, will perform better on Spark. If you want to actually train and fine-tune models (not just run them), the spark wins on compute alone.
These products are so niche and financially doesn't even make sense when you can probably buy a decade of cloud AI that will continue to advance at a rapid pace, over buying any of these solutions for local LLM, imo. Unless you're generating things that clouds would stop you doing... those peoples.
The DGX spark actually made sense for network engineers to dev time on it locally and knew it was the same output for the datacenters. Not sure who will spend that kind of cash on that laptop.
A different PoV is Apple is the best CPU. Nvidia is the best GPU. No one else is on the level of either in technical competence except Huawei. But that firm is sanctioned to high hell.
Wilhelm_85
Gold Member
Topher
Identifies as young
Bandwidth ain't everything when even the M3 ultra had ~819 GB/s 512 GB
It's good for generation time. Apple M's are weak on compute side.
DGX wins in raw compute and FP4 allowing large models. It's prefill performances likely not beaten until M6, if even that.
- Prefill → high compute device.
- Decode → high memory-bandwidth device.
Good benchmarking and explanation here.
Combining NVIDIA DGX Spark + Apple Mac Studio for 4x Faster LLM Inference with EXO 1.0
Disaggregating Prefill and Decode: Faster First Tokens, Faster Streamsblog.exolabs.net
So they are apples and oranges.
NVIDIA DGX Spark vs Mac Studio M5 Max for Local AI 2026 | Presenc AI
Head-to-head 2026 comparison of NVIDIA DGX Spark and Mac Studio M5 Max 128GB for local LLM inference, fine-tuning, and on-device RAG: throughput, memory...
A complex agentic coding tasks where the agent has a big system prompt, reads in tool inputs and files, will perform better on Spark. If you want to actually train and fine-tune models (not just run them), the spark wins on compute alone.
These products are so niche and financially doesn't even make sense when you can probably buy a decade of cloud AI that will continue to advance at a rapid pace, over buying any of these solutions for local LLM, imo. Unless you're generating things that clouds would stop you doing... those peoples.
The DGX spark actually made sense for network engineers to dev time on it locally and knew it was the same output for the datacenters. Not sure who will spend that kind of cash on that laptop.
Nice. Thanks. I'm going to dive into all that later
Bry0
Member
Bandwidth ain't everything when even the M3 ultra had ~819 GB/s 512 GB
It's good for generation time. Apple M's are weak on compute side.
DGX wins in raw compute and FP4 allowing large models. It's prefill performances likely not beaten until M6, if even that.
- Prefill → high compute device.
- Decode → high memory-bandwidth device.
Good benchmarking and explanation here.
Combining NVIDIA DGX Spark + Apple Mac Studio for 4x Faster LLM Inference with EXO 1.0
Disaggregating Prefill and Decode: Faster First Tokens, Faster Streamsblog.exolabs.net
So they are apples and oranges.
NVIDIA DGX Spark vs Mac Studio M5 Max for Local AI 2026 | Presenc AI
Head-to-head 2026 comparison of NVIDIA DGX Spark and Mac Studio M5 Max 128GB for local LLM inference, fine-tuning, and on-device RAG: throughput, memory...
A complex agentic coding tasks where the agent has a big system prompt, reads in tool inputs and files, will perform better on Spark. If you want to actually train and fine-tune models (not just run them), the spark wins on compute alone.
These products are so niche and financially doesn't even make sense when you can probably buy a decade of cloud AI that will continue to advance at a rapid pace, over buying any of these solutions for local LLM, imo. Unless you're generating things that clouds would stop you doing... those peoples.
The DGX spark actually made sense for network engineers to dev time on it locally and knew it was the same output for the datacenters. Not sure who will spend that kind of cash on that laptop.
What's a Mac Studio m5 max? That doesn't exist.
IntentionalPun
Ask me about my wife's perfect butthole
AI slop article pretending like the next Mac Studio is already out, and probably using the specs of the M5 Max MacBook that came out in March as a comparison.
Or it magically knows the price and specs of a device that hasn't been announced yet because it's AI obviously!
Last edited:
Insane Metal
Member
Lol fuck you Nvidia.
YCoCg
Member
You realise how sad that makes you sound?
Buggy Loop
Gold Member
But he's right
Yelling at clouds or just being part of the anti-AI church is not gonna stop the development of this and if you just resist changes, you'll be left behind. I don't think anything he said was even controversial.
The genie is not going back into the bottle.
RoboFu
One of the green rats
I love AI. it's fucking up a lot of code base which is a good side business for me!
Why do you pro AI guys always talk as if "LEARNING" AI is thing? thats the whole promise of AI. You do not have to have any skill or knowledge to do/create something. Hell there's a middle age man with a viral AI created song now. took him a whole 15 seconds to type in the prompt.
Why do you pro AI guys always talk as if "LEARNING" AI is thing? thats the whole promise of AI. You do not have to have any skill or knowledge to do/create something. Hell there's a middle age man with a viral AI created song now. took him a whole 15 seconds to type in the prompt.
ResurrectedContrarian
Suffers with mild autism
Haven't looked at all the details of this yes, but -- having lived in the wonderful Apple Silicon world for a couple of years, it really is hard to imagine any laptop being useful or taken seriously without unified memory in the future. The fact that I can just load 30GB+ transformer models (and with a larger Max series mac, much higher than that!) and just sample tokens from it fast at will, it changes everything in terms of the kinds of things you just do now entirely on local machine... image generation, even video generation, all kinds of autonomous local agents, voice recognition and cloning at speed, etc.
PauloRoberto
Member
I love how some idiots think that AI is a thing that will never go away and will never change... we are ALREADY seeing lots of 'AI' initializes scaling down because of cost and efficiency... it looks like no one remember lots of tech things that appeared and went away... and yes, AI, as it stands today, IS ONE OF THOSE THINGS. It's too error prone, too costly and too tied to specific vendors in almost every commercial case. And the bills started to arrive.
Topher
Identifies as young
Not true. I created a very detailed markdown file that AI uses as a guide when I have it write code. So while I have it working on one module in an app, I can be working on something else. The markdown file details the architecture and patterns being used in the app. It also tells AI to document, write test plans, and unit tests along the way. So if you think I can just tell it to "write code" like some guy taking two minutes to have AI create a song the you really do not have a very good grasp of what is involved.
But you are right. It will fuck up a code base unless you learn how to use it as a tool properly.
Last edited:
Tobimacoss
Member
Microsoft and NVIDIA’s Surface Laptop Ultra pushes Windows on Arm into high‑performance territory
Microsoft and NVIDIA unveil the Surface Laptop Ultra, a 128GB RAM beast with Blackwell graphics and a mini-LED display that redefines performance for Windows on Arm.
www.windowscentral.com
A beast of a laptop, 128 gb ram. At least they used the term Ultra. If only they used it for new Gamepass tier.
Topher
Identifies as young
Microsoft and NVIDIA’s Surface Laptop Ultra pushes Windows on Arm into high‑performance territory
Microsoft and NVIDIA unveil the Surface Laptop Ultra, a 128GB RAM beast with Blackwell graphics and a mini-LED display that redefines performance for Windows on Arm.
A beast of a laptop, 128 gb ram. At least they used the term Ultra. If only they used it for new Gamepass tier.
If it is like other Surface laptops then it will be severely overpriced compared to the competition. Either way, all of the Nvidia laptops offer "up to" 128GB.
Last edited:
YCoCg
Member
Everyone will be left behind, it's going to put people out of jobs, raise the price of hardware even more, have it be so only a handful of companies control our society, etc. It's the "you'll own nothing and be happy" playback but you lot are cheering it on because it's "the future", when will you realise this isn't some utopian scientists here, the goal here is to screw everyone over and make as much money as possible whilst removing the average persons choice.
If things keep going down this path then by 2040 people won't even be able to own pc hardware anymore, we'll just be sold AI terminals that connect to the cloud, no local storage.
Topher
Identifies as young
That sounds like lines we've been fed by AI corporation CEOs trying to sell their souls to stockholders. I just don't see it man.
{insert code here}
Member
"The PC is being reinvented," said Jensen Huang, founder and CEO of NVIDIA. "For forty years, you launched apps. Click. Type. With RTX Spark and Microsoft Windows, you ask — and the PC does the work. RTX Spark brings everything NVIDIA has built — CUDA, RTX, our AI platform — into a single superchip. Local agents. Frontier models. Creative workflows. RTX games. All on a laptop. This is the new PC. The personal AI computer."
Last edited:
Autistic_Pancake
Member
> RTX Spark
Basically the DGX: mobile and DGX: electric boogalo.
Brb, selling my kidney to run dense models like Qwen 3.6 27B / Gemma 4 31B at miserable speed with no tensor parallelism, quantized into 4-bit lobotomites.
inb4 GAF begins to defend it once again, like when someone replied in one of those threads - "you can't stack 3090s cause that's not how LLMs work" - man, that was just...
Is THIS the future of personal computing the people really want? I know some claim it's useful for "CUDA developers", but why the rest must suffer?
IF the cost wasn't so high, then yes, perhaps it wouldn't seem so crazy to buy such a device. Realistically, though, it's a step toward herding us into the cloud computing prison.
Basically the DGX: mobile and DGX: electric boogalo.
Brb, selling my kidney to run dense models like Qwen 3.6 27B / Gemma 4 31B at miserable speed with no tensor parallelism, quantized into 4-bit lobotomites.
inb4 GAF begins to defend it once again, like when someone replied in one of those threads - "you can't stack 3090s cause that's not how LLMs work" - man, that was just...
Is THIS the future of personal computing the people really want? I know some claim it's useful for "CUDA developers", but why the rest must suffer?
IF the cost wasn't so high, then yes, perhaps it wouldn't seem so crazy to buy such a device. Realistically, though, it's a step toward herding us into the cloud computing prison.
ResurrectedContrarian
Suffers with mild autism
not really the right comparison for LLMs though... going forward even 32GB won't be enough to fit the better models. If pure GPUs can't get us to 80GB+ models in the future on home hardware, then they really are on the path to being obsolete for local LLM / token sampling purposes, given that you actually can fit models of that size or larger on these unified memory / ARM chips from Apple or now NVIDIA. At that point, buying 4 RTX gets pretty extreme in cost compared to a single M-series Max or DGX next version or whatever is coming down the pipeline getting cheaper year after year, yes?> RTX Spark
Basically the DGX: mobile and DGX: electric boogalo.
Brb, selling my kidney to run dense models like Qwen 3.6 27B / Gemma 4 31B at miserable speed with no tensor parallelism, quantized into 4-bit lobotomites.
inb4 GAF begins to defend it once again, like when someone replied in one of those threads - "you can't stack 3090s cause that's not how LLMs work" - man, that was just...
Is THIS the future of personal computing the people really want? I know some claim it's useful for "CUDA developers", but why the rest must suffer?
IF the cost wasn't so high, then yes, perhaps it wouldn't seem so crazy to buy such a device. Realistically, though, it's a step toward herding us into the cloud computing prison.
(But for what it's worth, I would recommend a "true" GPU if doing image & video generation. Different trade-off from LLMs, more raw compute constrained than model size.)
Last edited:
Autistic_Pancake
Member
Yes, I know it's a valid point. The issue is that it's only 128GB. M3 Ultra, for example (when its price had not skyrocketed yet) allowed to get 512GB which truly enabled to run some huge LLMs.
But this little thing is just a misery-box I'd personally never prefer over 4x RTX 3090 running undervolted at 70% power limit, in a system with 128GB DDR5 (if bought before the price crisis). Small-to-medium dense models would fit entirely in VRAM, running at FP16 doing agentic shit and whatnot. Large MoE models - no problems either, split between VRAM / RAM, once again at better speeds (not to mention 96+128 would allow bigger models than just 128). I guess the only real caveat is power consumption, or rather, your bills. And that of course totally depends on where you live.
ResurrectedContrarian
Suffers with mild autism
for what it's worth, I'm on Apple Silicon machines because that's what I can reliably get my employers to 100% expense, but I'd probably trade for 4x RTX GPUs if given the option ... my image and video gen would finally fly, it's painful right now
S0ULZB0URNE
Member
Buggy Loop
Gold Member
Socrates and many others thinking that the invention of writing would ruin human intelligence (~3000 BCE)
Printing press, it will lead to information overload, urging to stop using them as the public would suffer from confusing and harmful abundance of books. (1600s)
Skilled craftsmen and weavers will lose their jobs from the automation of the machines in the industrial revolution! (early 1800's)
Trains traveling faster than 30 miles per hour would cause the human body to melt (1800s)
Light bulbs causes blindness and made women infertile! (late 1800s)
Telephones will kill all real face-to-face conversation and ruin society's social fabric! (late 1800s)
Internet
emails hurt IQ more than pot - CNN
The vanishing of secretariat pools with computers and anxiety as the rest of the workforce had to fucking type
Now it's AI.
By 2040, in a fucking long ass time, owning PC hardware and no local storage, that's really what's this about? Are you seriously thinking it would not have went that direction without AI ? Peoples have speculated on this way before AI.
Last edited:
YCoCg
Member
Socrates and many others thinking that the invention of writing would ruin human intelligence (~3000 BCE)
Printing press, it will lead to information overload, urging to stop using them as the public would suffer from confusing and harmful abundance of books. (1600s)
Skilled craftsmen and weavers will lose their jobs from the automation of the machines in the industrial revolution! (early 1800's)
Trains traveling faster than 30 miles per hour would cause the human body to melt (1800s)
Light bulbs causes blindness and made women infertile! (late 1800s)
Telephones will kill all real face-to-face conversation and ruin society's social fabric! (late 1800s)
Internet
emails hurt IQ more than pot - CNN
The vanishing of secretariat pools with computers and anxiety as the rest of the workforce had to fucking type
Now it's AI.
By 2040, in a fucking long ass time, owning PC hardware and no local storage, that's really what's this about? Are you seriously thinking it would not have went that direction without AI ? Peoples have speculated on this way before AI.
Yeah most of those still have humans behind them though, yet here we're supposed to trust daddy Musk, Altman, Jensen, etc and a bunch of all these companies at the top that we should just give into AI and let it do everything for us.
No doubt on our rented devices that we pay a monthly subscription for oh and because of all the bullshit government worrying about "children", we'll need to have our official government digital ID linked to these devices so we can be monitored, oh wait, I mean protecting the children, whilst we also let AI teach our children as we watch our AI curated movies and shows and listen to our AI made up music that we can't talk to other people about because we all get tailored made stuff for our own little bubbles.
peish
Member
Yes we can stopped that from happening! By working together to stamp out this AI bubble!
Besides why would I need a computer to generate tokens so that it can then automate some stuff??? I just want to click and do it myself! Why would anyone want waste hardware on a middleman?
TheThreadsThatBindUs
Member
Not if we can't afford the parts.
peish
Member
I dont see the point of local AI if you are not a coder? You dont need to memorise all the syntax and contexts for sql…
But for a normal windowser, any point in speaking to your pc, "windows, check neogaf, if any new thread on project helix, post a few concerned trolling replies in the span of 4 hours, tone is dry wit but with some technology terminologies "
But for a normal windowser, any point in speaking to your pc, "windows, check neogaf, if any new thread on project helix, post a few concerned trolling replies in the span of 4 hours, tone is dry wit but with some technology terminologies "
WhamBamThankYouMaam
Member
1. Nvidia etc ensures memory prices increase massively.
2. Nvidia shows new mschine that is fantastic with 128 gb ram.
3. No one can afford it.
4. Everyone is stuck with same old local models with low amounts of ram.
2. Nvidia shows new mschine that is fantastic with 128 gb ram.
3. No one can afford it.
4. Everyone is stuck with same old local models with low amounts of ram.
Badlucktroll
Member
soon pc will become WISH machine
StereoVsn
Gold Member
The issue is that if you don't already have this setup, it's prohibitively expensive right now. And of course 3090s would all be used with no warranty.
Might as well wait for M5 Studio and buy that at say Microcenter with a slight discount.
Gameplay Gods Bless
Member
DGX is always more expensive than the RTX equivalent.
Pedro Motta
Gold Member
It's same chip tho? What is the difference? SOC's on display at Computex had a date of fabrication of 2024, it's the exact same chip. This is only arriving now because Microsoft and Windows weren't ready before. Macbook Pro with 128Gb of unified memory are also in the 5K range.
Topher
Identifies as young
What has been the RTX equivalent before this?
Dr. Samuel Hayden
Member
I heckin love slop!
MiguelItUp
Member
Not a fan of any of this, personally. AI can be a great tool, I know this and I've used it. I'm not one of those radical haters. But I am over the obsession of it all, but I also know that it's unfortunately not stopping anytime soon. The revenue Nvidia has been obtaining from Data Centers, Hyperscale, and AI Clouds (Industrial & Enterprise) has been insane. Especially compared to Edge Computing which now houses PCs and consoles, you know the stuff we care about.
I just want the bubble to pop.
I just want the bubble to pop.
MCH2024
Member
It's a 2024 ARM Super Core and Windows is a shitty OS that nerfs GB6 scores.
Nevertheless yea it's weaker than 2025/2026 ARM Super Cores.
It's still considerably and vastly superior to X86. Let's not get things mixed up.
And the selling point is the REVOLUTIONARY GPU.