Microsoft and Intel nail down what an "AI PC" is

Daniel Sims

Posts: 1,375   +43
Staff
Why it matters: Microsoft and multiple chipmakers have spent months heralding the arrival of the "AI PC," which utilizes generative AI and large language models to facilitate various tasks in new ways. However, the definition of an AI PC remains somewhat unclear. At a recent event in Taipei, Intel began defining the specifics that it agreed upon with Microsoft.

At Intel's recent AI summit in Taipei, the company informed the public that Windows PCs will soon leverage NPUs (Neural Processing Units) to run Microsoft Copilot assistant locally. This is the first confirmation of rumors that emerged regarding Copilot and increasingly prevalent hardware-based accelerators late last year.

Microsoft began pushing Copilot to Windows 11 users last year, but that currently operates in the cloud. The company is expected to enhance the AI assistant by integrating NPUs starting with the 24H2 update, possibly arriving this summer.

While Intel didn't provide a specific date for the upgrade, a spokesperson mentioned that Copilot will begin handling many, but not all, tasks locally using NPUs, reducing its reliance on the cloud.

Microsoft is seeking to inaugurate the new "AI PC" era by introducing a dedicated Copilot key to the specifications for Windows keyboards, and Intel is collaborating by incorporating the key into its official requirements for an AI PC.

Previously, the minimum AI PC specs were defined by the presence of a CPU, GPU, and NPU. Additionally, the company has listed 40 TOPs of NPU performance as a requirement for next-generation AI PCs.

NPUs have begun to appear in Intel's Meteor Lake (a.k.a. Core Ultra), AMD's Strix Point, and Qualcomm's Snapdragon X Elite platform.

Recent graphics cards from Nvidia, Intel, and AMD can also locally run AI processes like Nvidia's RTX-powered chatbot. However, Intel stated that NPUs are designed to alleviate the extra workload on CPUs and GPUs, and testing has shown to improve battery life. Nevertheless, AI applications can also consume significant amounts of memory, and it remains unclear how this might impact AI PC system requirements.

In Taipei, Intel also announced a new initiative to provide developers with resources to create new software to leverage AI. Currently, AI applications mostly involve text and image generation, image processing, and enhanced search engines.

Recent Windows Insider builds indicate that Microsoft is testing features that might enable Copilot to automatically navigate settings, find items based on text descriptions, or perform other tasks to expedite workflows.

Permalink to story.

 
So Microsoft would like me to buy a new computer so that the service that they currently host in the cloud at their expense, will now run on my local hardware instead? And shall we guess that despite it running on local hardware that all requests made will still be logged back to Microsoft for "quality control"?

I'm so grateful Microsoft is providing me this opportunity to contribute.
 
The keyboard thing is kinda new, and will take some getting used to. Not sure I like it. Has the biggest impact on laptops, not so much on desktops.

That said, the NPU thing kinda makes sense to me now: it frees up the GPU and CPU and is more energy efficient at AI tasks (because it is designed for that specifically, remains to be seen how much more efficient). Still doesn't solve the memory issue.

Honestly, I'm surprised not to see a memory requirement for AI PCs. If you thought TPM for Windows 11 was bad, just imagine "oh by the way you need 32 Gigs of RAM for Windows 12 so we can run copilot on your device, but we recommend 64 Gigs of RAM for an optimal experience" (a quantized 7B model and a bit of conversation can fit in 8 Gigs of VRAM/RAM just fine, assuming you don't need that RAM for anything else, but 7b models are tiny and error prone).

Windows 12 apparently isn't happening (this year; it's just a Win 11 update), so no new hardware requirements from that side. They still could've proposed a RAM requirement for AI PCs though. Probably would exclude too many so they didn't bother. I still don't know how Microsoft expects inference to happen on your device locally with limited RAM and still be high quality unless copilot is built on a tiny model. Maybe they'll keep the enhanced version behind a subscription, but somehow I doubt it, except for Office 365.
 
Microsoft announces requirements for a PC I will never own.
So Microsoft would like me to buy a new computer so that the service that they currently host in the cloud at their expense, will now run on my local hardware instead? And shall we guess that despite it running on local hardware that all requests made will still be logged back to Microsoft for "quality control"?

I'm so grateful Microsoft is providing me this opportunity to contribute.
You forgot requiring your account to be a MS account linked to everything else in your life, to ensure "positive behavior".
 
Microsoft and Intel nail down what an "AI PC" is ... really? I'll tell you exactly what an AI PC is; it's a marketing gimmick to sell you more hardware you don't actually need. PC sales have been slumping, so they needed a new must-have gimmick they can push on everybody and try to increase sales.

The next step will be AI phones, because those sales are cratering too. People need a reason to keep buying thousand dollar phones year after year and tiny incremental changes are no longer cutting it. Next year you must buy an AI phone or you'll be falling behind!

PROTIP: There's no such thing as an AI PC, in fact there's no AI hardware of any kind anywhere in the world. It's just a catch all moniker to make things seem more advanced than they actually are. ChatGPT and the other nonsense bots are not AI. Someone famous (I can't remember who) said ChatGPT is just a glorified tape recorder and they're exactly right.
 
I see no point in them nailing down what is an "AI PC". Like most consumer can't care less about AI, and its all marketing fluff. Ultimately, it is not like consumers have a choice here. When you need to get a new laptop for instance, its going to have all these AI hardware/ NPUs built in them whether you like it or not.

And I agree that the PC/ mobile market needs something that they can hype up to boost sales, and AI is the marketing they are using now.
 
The keyboard thing is kinda new, and will take some getting used to. Not sure I like it. Has the biggest impact on laptops, not so much on desktops.

That said, the NPU thing kinda makes sense to me now: it frees up the GPU and CPU and is more energy efficient at AI tasks (because it is designed for that specifically, remains to be seen how much more efficient). Still doesn't solve the memory issue.

Honestly, I'm surprised not to see a memory requirement for AI PCs. If you thought TPM for Windows 11 was bad, just imagine "oh by the way you need 32 Gigs of RAM for Windows 12 so we can run copilot on your device, but we recommend 64 Gigs of RAM for an optimal experience" (a quantized 7B model and a bit of conversation can fit in 8 Gigs of VRAM/RAM just fine, assuming you don't need that RAM for anything else, but 7b models are tiny and error prone).

Windows 12 apparently isn't happening (this year; it's just a Win 11 update), so no new hardware requirements from that side. They still could've proposed a RAM requirement for AI PCs though. Probably would exclude too many so they didn't bother. I still don't know how Microsoft expects inference to happen on your device locally with limited RAM and still be high quality unless copilot is built on a tiny model. Maybe they'll keep the enhanced version behind a subscription, but somehow I doubt it, except for Office 365.
btw, the tpm 2.0 requirement of windows 11 is because the previous tpm 1.2 only specifies sha1 hashing which is no longer considered as secure hash
 
The average gamer only cares abt how many extra FPS AI can give him.

If it's zero FPS you can imagine that interest will be low.
 
Honestly, I'm surprised not to see a memory requirement for AI PCs. If you thought TPM for Windows 11 was bad, just imagine "oh by the way you need 32 Gigs of RAM for Windows 12 so we can run copilot on your device, but we recommend 64 Gigs of RAM for an optimal experience" (a quantized 7B model and a bit of conversation can fit in 8 Gigs of VRAM/RAM just fine, assuming you don't need that RAM for anything else, but 7b models are tiny and error prone).

16GB was previously mentioned, I'm not sure if that requirement was removed: https://www.techpowerup.com/317978/...quirement-for-copilot-and-windows-ai-features
 
Still a sollution looking for a problem…

And I’m sure they just rebranded an existing compute unit inside the embedded GPU to NPU…
 
So I now know WHAT to look for when I want to save a few bucks (and will NOT be running Windows on my system anyway.) If it says it's an "AI PC", keep looking (similar to a few years ago when they started pushing touch screens -- hard -- and I found a good price for some notebook just because it had a conventional screen on it.) I think the vendors have moved away from this more recently, realizing people preferred a low price over the $50-100 extra spend for a touch screen (at the low end) and (at the higher end) may not worry as much about that price but would rather the $50-100 went towards a slightly faster CPU or GPU, more RAM, and more SSD space, or just plain better build quality, instead.
 
Back