Will NPUs soon become useless silicon? — Windows 11’s Local AI is getting GPU support

Is Microsoft making NPUs useless? – Windows 11’s Local AI is getting GPU support​

Microsoft is giving Windows 11’s Local Language AI Models official Nvidia GPU support

With Microsoft’s Windows App SDK 2.2 Experimental 9 update, as spotted by Windows Latest, Windows 11’s language model APIs will run on non-Copilot+ PCs with specific GPUs. Right now, this is limited to Nvidia RTX 30 series GPUs (or newer) with at least 6GB of VRAM. In time, this support could come to AMD and Intel GPUs.

https://overclock3d.net/news/softwa...-windows-11s-local-ai-is-getting-gpu-support/

I mean it makes sense that they'd allow this, but at the same time it doesn't necessarily make the NPUs useless, they still have their areas to be used in and can tak advantage of not needing the higher power usage of the gpus for it.

Balkroth said:

I mean it makes sense that they'd allow this, but at the same time it doesn't necessarily make the NPUs useless, they still have their areas to be used in and can tak advantage of not needing the higher power usage of the gpus for it.

Click to expand...

the way I see it Intel & AMD can add new instructions to CPU to handle low throughput high latency NPU use-cases

nvidia will push for GPU to handle as much as possible

maybe the memory access gives NPU an leg up over GPU. But unified memory & directX changes should be able to address that

Intel already has Xe Matrix Extensions. It's used for video upscaling. Intel also has the Arc Pro dedicated NPUs.
Nvidia is moving away from GPUs. AFAIK no new Nvidia GPUs till 2028.

LLMs with good performance require 100+GB of memory, which makes local LLMs more or less useless in the consumer market. Small and crappy models can run fine on the CPU.

I feel really not, from groq, cerebras to other like HFboard kyle project, specialized for AI inference asic ("NPU"), have probably a future.

Apple-google will continue to make them and we could see them become quite common on windows Laptop of the future, next generation or the future one, Nvidia GPU will come with on package npu will be my guess, groq chiplet LPU with sram weight, making them always there, on APU soc and on discrete gpu has well.

Would not be surprising if we see an exact opposite, 100% of inference involving for at least some part of it a silicon that was built from the ground up only for it and doing only that.

One good reason for it, those specialized for inference only silicon of the sram and other strategy to have the weight directly on the compute will use older TSMC node really well

LukeTbk said:

I feel really not, from groq, cerebras to other like HFboard kyle project, specialized for AI inference asic ("NPU"), have probably a future.

Apple-google will continue to make them and we could see them become quite common on windows Laptop of the future, next generation or the future one, Nvidia GPU will come with on package npu will be my guess, groq chiplet LPU with sram weight, making them always there, on APU soc and on discrete gpu has well.

Would not be surprising if we see an exact opposite, 100% of inference involving for at least some part of it a silicon that was built from the ground up only for it and doing only that.

One good reason for it, those specialized for inference only silicon of the sram and other strategy to have the weight directly on the compute will use older TSMC node really well

Click to expand...

how much die space would you allocate for the NPU ?

On zen 6 desktop CPU, AMD has replaced the igpu on the iod die with an NPU
Is that kind of trade-off worth it in your opinion ?

stuffmonger said:

LLMs with good performance require 100+GB of memory, which makes local LLMs more or less useless in the consumer market. Small and crappy models can run fine on the CPU.

Click to expand...

But lot of inference being made are not LLM has you say, upscaling (video or 3d) use small model, the upcoming texture and other neural base asset decompression, specially the NVFP4 (4-bit) version of them they will able to use.

And that 100gb memory it depend how much context you need, latest gemma 4 diffusion, 26Billion MoE model does not require that level of memory to be fast and quite good.... and importantly people will have 128-512-1024gb of memory in their machine if it is worth it.

LukeTbk said:

But lot of inference being made are not LLM has you say, upscaling (video or 3d) use small model, the upcoming texture and other neural base asset decompression, specially the NVFP4 (4-bit) version of them they will able to use.

And that 100gb memory it depend how much context you need, latest gemma 4 diffusion, 26Billion MoE model does not require that level of memory to be fast and quite good.... and importantly people will have 128-512-1024gb of memory in their machine if it is worth it.

Click to expand...

the only 2 use cases I can think of for NPU on consumer PCs

1. spellcheck/auto correct
2. to spy on you (for the govt)

Marees said:

how much die space would you allocate for the NPU ?

Click to expand...

If it is on older much cheaper TSMC and packaging get good enough a lot, but it will be a different die in a chiplet design.

DLSS weight take about 120 to 160mb, say by shifting to nvfp4 they achieve to stay about the same size despite couple of generation bump in complexity (dividing their size by 2x-4x going down in bitsize to compensate), they could need on what would be quite old and mature TSMC N6 90mm for the sram+compute, if RR and texture run there too, maybe a 120mm type as packaging get better and better, chiplet could become the norm.

This would give them "almost free" dlss-texture generation and others from the rest of the gpu and the ability to generate a lot of frames per seconds. Game engine gain tensor core back to make more custom neural affair.

Marees said:

the only 2 use cases I can think of for NPU on consumer PCs

1. spellcheck/auto correct
2. to spy on you (for the govt)

Click to expand...

DLSS type (like microsoft use them) being an other obvious one, voice-video call treatment and all other things phone have been using them for years and anything a GPU could do without them but better. Local work agents and local LLM.

NPU is just a term to say an ASIC for ai inference, if that workload become popular enough, that what will be used like google TPU and others to do it.

stuffmonger said:

LLMs with good performance require 100+GB of memory, which makes local LLMs more or less useless in the consumer market. Small and crappy models can run fine on the CPU.

Click to expand...

Is 100 GB+ used by ChatGPT or Claude Opus? For privacy I may want to run a local LLM, even it is not as "large." Over time i expect lots of local models, trained on a user's own data.

philb2 said:

Is 100 GB+ used by ChatGPT or Claude Opus? For privacy I may want to run a local LLM, even it is not as "large." Over time i expect lots of local models, trained on a user's own data.

Click to expand...

those are secrets, but for the most expensive version of those model, it is estimated that it is counted in terrabyte of memory for trillions of parameter, maybe 3-5 for Mythos.

128gb of ram that more for a 70B version of deepseek at 8Bits type of affair, mix of expert they call (only a subset of the weight are used at the same time even if the whole model is loaded) can make large memory but slow memory like apple M, strix Halo, nvidia digits has work at reasonable speed, even with large total sized model.

LukeTbk said:

those are secrets, but for the most expensive version of those model, it is estimated that it is counted in terrabyte of memory for trillions of parameter, maybe 3-5 for Mythos.

128gb of ram that more for a 70B version of deepseek at 8Bits type of affair, mix of expert they call (only a subset of the weight are used at the same time even if the whole model is loaded) can make large memory but slow memory like apple M, strix Halo, nvidia digits has work at reasonable speed, even with large total sized model.

Click to expand...

If local large models can use unified models, I would have expected a lot of 64 GB and 128 GB memory kits to be sold. But that won't happen at current DRAM prices.

philb2 said:

If local large models can use unified models, I would have expected a lot of 64 GB and 128 GB memory kits to be sold. But that won't happen at current DRAM prices.

Click to expand...

I am not sure what unified model mean ? you mean multimodal with sound and image ?

memory kits are often a bit slow memory for this use (still imagine they sold everything they could do), what they sold a lot is high vram gpu and apple M3 ultra type and they are selling a lot of them even with high price, because of how much work people expect to do with them.

High price do often mean a lot are being sold and the price is high because demand is high with a lot of buyers.

LukeTbk said:

I am not sure what unified model mean ? you mean multimodal with sound and image ?

memory kits are often a bit slow memory for this use (still imagine they sold everything they could do), what they sold a lot is high vram gpu and apple M3 ultra type and they are selling a lot of them even with high price, because of how much work people expect to do with them.

High price do often mean a lot are being sold and the price is high because demand is high with a lot of buyers.

Click to expand...

If unifiied memory local models do prove out then there will be a lot of VRAM demand, boosting pricing. But pricing is already painfully high due to the AI bubble.

Once people really start figuring out what AI can do then they're going to be significantly less inclined to trust these services much less plug their entire lives into them. This includes any local AI trojan horse that Microshit might come up with.

This shouldn't shock anyone. Especially since AI isn't a part of the OS, let alone a big part. CPU's have extensions for matrix multiplications and GPU's are getting better at it. Companies want you to use the cloud for AI as well.

LukeTbk said:

NVFP4

Click to expand...

That's only for Blackwell boards? None of those are for the consumer market.

LukeTbk said:

depend how much context you need, latest gemma 4 diffusion, 26Billion

Click to expand...

Unquantized (16 bit) this requires 60GB. Even with 4 bits, it's 15GB.
I've tested unquantized DeepSeek a year ago, and the results were obviously superior to quantized for a small number of NLP tasks.

LukeTbk said:

people will have 128-512-1024gb of memory in their machine if it is worth it

Click to expand...

It is not worth it for the vast majority of consumers. They're gonna sell their kids to slavery to generate meme videos?

philb2 said:

Is 100 GB+ used by ChatGPT or Claude Opus? For privacy I may want to run a local LLM, even it is not as "large." Over time i expect lots of local models, trained on a user's own data.

Click to expand...

Very likely so, since they use Nvidia H200 for inference, which has 140 something GB.


As it stands, the only viable consumer application for local models is upscaling, which is ironically cannibalized by Nvidia limiting VRAM for new consumer GPUs.

stuffmonger said:

That's only for Blackwell boards? None of those are for the consumer market.

Click to expand...

From the RTX 5050 discrete gpu to the many blackwell laptop (and the upcoming N1X), NVFP4 support is common and will be even more in 3 years in the newly bought consummer hardware, so will native FP4 support among other sellers (int4 on apple already there).

stuffmonger said:

Unquantized (16 bit) this requires 60GB. Even with 4 bits, it's 15GB.

Click to expand...

yes, and the impact of quantisation is getting lower and lower.

stuffmonger said:

It is not worth it for the vast majority of consumers. They're gonna sell their kids to slavery to generate meme videos?

Click to expand...

Consumer of this (expensive PC) are the company buying the work computer of people that work on computer, employer (or with their tax expense independant company) will be what pay, it will still be much cheaper than an used car, it will be aobut what computers did cost in 1994 adjusted in today dollar for the very large memory and very capable version, which people that worker on computer were paying at the time. For the majority of consummers outside work for compute, it will be smarthphone and gaming PC.

Many will be too young to remember, but people were paying a fortune for computers, with which you could not do much with versus today, people working in SFX, solidworks and other field are still paying a lot for them.

stuffmonger said:

Very likely so, since they use Nvidia H200 for inference, which has 140 something GB.

Click to expand...

Inference for those giant model are not made on a single gpu, but a rack that as unified memory pool, an NVL72 rack has 13.5 TB of HBM3e.

stuffmonger said:

As it stands, the only viable consumer application for local models is upscaling, which is ironically cannibalized by Nvidia limiting VRAM for new consumer GPUs.

Click to expand...

To a point vram usage of those small DLSS type model (120-150-160MB) tend to be smaller than the gain in vram made by rendering at lower resolution, if anything low vram because of current cost could be why the next generation of console will use upscaling so much, same for ai compression-decompression of asset.

DukenukemX said:

This shouldn't shock anyone. Especially since AI isn't a part of the OS, let alone a big part.

Click to expand...

Not sure how relevant AI being part of the os or not is there and what should not shock anyone that microsoft AI api access GPU... of course ?

That the 40-50 tops npu of 2023-2024 will never have been useful.. of course. But that completlety different than npu becoming useless as a whole, cerebras-groq like NPU able to run current 120billion model at over 500 token per second are a complete different animal.

LukeTbk said:

SFX, solidworks

Click to expand...

That's enterprise customers. Tiny market. SFX is contracting to correct over-hiring that happened 5 years ago.
I give no shits about enterprise. Maybe you have a company that benefits from buying $30K boards, then I'm a total buffoon, and if not, your statements have no weight.

N1X is gonna be $5,000? With 25W idle, it's gonna last last only a few hours vs cheap ass $1,000 macbook air that lasts for 20 hours.
So spending $4,000 extra, and sacrificing battery life, will save me $20/mo meant for ChatGPT Plus?
It reminds me of Light L16, which was $2,000 when it was released, then flopped. Now it's $200.
I don't mind waiting for 10 years to grab an N1X for $500.

LukeTbk said:

To a point vram usage of those small DLSS type model (120-150-160MB) tend to be smaller than the gain in vram made by rendering at lower resolution, if anything low vram because of current cost could be why the next generation of console will use upscaling so much, same for ai compression-decompression of asset.

Click to expand...

The problem is that 8GB is not enough on its own, so even if DLSS took 0 VRAM, it'd still be a problem.
Nvidia would have released 16GB budget GPUs if there were no memory shortage, and if gaming were still the main profit maker for them.

fine by me. I would still like to have the option to run AI offline. I hope the prices of GPUS dont go up because of this...who am i kidding.

stuffmonger said:

That's enterprise customers.

Click to expand...

Yes that who tend to be buying new PC outside gamers, who think that expensive working local AI agentic is for ? If not people that work on computers.

If $20 a month ChatGPT work for someone and have no local data laws-regulation in their field going on that limit that kind of use, they have no use for edge AI usually.

People that work on computers is not a niche market(that a giant proportion of the world workforce in the western world that what 33-45% of people ?), Solidworks workstation is just an example that when an expensive computer give an edge to their employee over non-expensive one, company do buy them, it was just that for many things a chromebook was powerful enough, MacBook air for bigger user, that why they did not bought expensive stuff. We can look at how much they spent on workchair and desk to have a clue.

LukeTbk said:

Not sure how relevant AI being part of the os or not is there and what should not shock anyone that microsoft AI api access GPU... of course ?

Click to expand...

It was Microsoft who pushed for NPU's because of Copilot and Recall. NPU's did exist before but we also never made use of them before.

That the 40-50 tops npu of 2023-2024 will never have been useful.. of course. But that completlety different than npu becoming useless as a whole, cerebras-groq like NPU able to run current 120billion model at over 500 token per second are a complete different animal.

Click to expand...

The whole point of an NPU is to do AI tasks in the background while not consuming CPU/GPU resources and a lot of power. Microsoft just realized that GPU's do it better, and you will likely rarely use AI on your Windows machine, which the GPU is perfect. The reality is that people rarely use AI and when they do it's ChatGPT and Gemini, which isn't using your NPU. There are NPU's in Google's and Microsoft's ARM based server chips. Even then, Google has TPU's while Nvidia has Tensor Cores. So yea, NPU's are worthless and the only reason the industry backed this was because Microsoft thought AI would change the world.

DukenukemX said:

The whole point of an NPU is to do AI tasks in the background while not consuming CPU/GPU resources and a lot of power.

Click to expand...

that not what google do with them or the one in the upcoming Rubin server rack, NPU just mean a silicon made specially for AI, it will be very efficiant at it, if you make it big it will also be very fast at it, much faster than non specialized hardware tend to be.

GPU do not do the inference part necessarily better than ASIC can, try this chatjimmy:
https://chatjimmy.ai/

Look how fast it can be (and that will be a huge * on the idea that local would be for latency purpose when it come to small data like text), is a gpu better than this Asic at running a small llama model ?

that what they are trying to do here:
https://hardforum.com/threads/welcome-to-archality-ai-inference-by-enthusiasts.2047929/

DukenukemX said:

So yea, NPU's are worthless

Click to expand...

Do you think the next generation of phone will stop having them, that in 2033 we will not have specialized silicon made for running part of the AI chain, like inference ?

DukenukemX said:

Even then, Google has TPU's while Nvidia has Tensor Cores.

Click to expand...

And the next generaiton of Nvidia datacenter (rubin) stack has NPU now, Groq LPUs a even narrower type of NPU, asic for inference is probably where it is going to be. Do not be surprised if Nvidia next discrete desktop GPU has NPU chiplet on cheaper older TSMC-Samsung-Intel nodes them, for graphical models, in less than 7 years from now.

Edge AI was far from being ready to be useful for agentic work 2 years ago and 40-50 tops type of device, not powerful enough for it, for conference call level of things and other smarthphone type of work thats been using NPU for years, it can work, but not much more than that, but that Copilot+ pc type of npu being out of date before local agentic AI workflow is polished enough for regular people to use is not a surprise (something i believed the moment thery did show up, that cloud would be cheaper, better) but a different subject all together to the idea that specialized made for AI silicon are useless and over...

Specialy basing this on some windows API gaining access to gpu work... as if the AI world is not running on Linux and Mac more.

LukeTbk said:

Do you think the next generation of phone will stop having them, that in 2033 we will not have specialized silicon made for running part of the AI chain, like inference ?

Click to expand...

You should ask me if CPU's will stop coming with NPU's, and the answer for both is no. The industry is stupid, so of course they'll continue to put NPU's. It'll take several CPU generations before manufacturers will admit that NPU's can be replaced with GPU's.

Specialy basing this on some windows API gaining access to gpu work... as if the AI world is not running on Linux and Mac more.

Click to expand...

Point is AI is rarely if ever used locally. I'm not aware of any optimizations for Windows that's making use of the NPU.

DukenukemX said:

Point is AI is rarely if ever used locally. I'm not aware of any optimizations for Windows that's making use of the NPU.

Click to expand...

Outside signal manipulation like DLSS, audio-video filtering type of use case of course (and other case when local data is big, making cloud less appealing than text base ai like LLM).

I am not sure what windows making use of NPU would be, but Adobe Photoshop and others can use NPU if present for their AI stuff. Apple having had them for a long time now and unified memory that make cpu-npus-gpu work on something using their ML framework in a bit transparent way will use them more.

DukenukemX said:

It'll take several CPU generations before manufacturers will admit that NPU's can be replaced with GPU's.

Click to expand...

The idea that Nvidia is safe making just GPU for inference, that no specialized ASIC for it (like Apple as had for a while) would ever show up seem just a lack of imagination.... the notion will be even stranger when GPU will have NPUs on them (like Apple latest gpus with their NPU inside them, as well as a 16 cores dedicated NPU on the soc), it is not if NPU can be replaced by gpu or not to do inference (gpu can be replaced by CPU to do it) it is which will be the best at doing it.

Outside a really big break on how things work, putting the model weight with the compute offer giant advantage (but limitation) that point toward we will continue to make specialized silicon just for AI inference and because those silicons does not scale well with smaller foundry node will be much cheaper to do per mm and will be a dedicated chips.

Some Nvidia RUBIN rack feature NPUs (even Nvidia do not replace them by GPU anymore...):


They are madde on some Samsung foundry node instead of using fancy new Tsmc node (SF4X), Nvidia local AI workstation of the near future could very well contain then (as chiplet on the discrete gpu or their own thing), 30-40x more inference per watt.

It can be used by itself to run very small model that will fit on them (like DLSS and some others non LLM are) or as a tool by gpu for big models (they can run a smaller heavy quantized version of a expert agents and make many thousand of free guess per seconds), many others project with many others strategy to make ASIC for ai inference are going, declaring none will work or none will be used on PC (just on other devices and servers), ok maybe...

But do not base this on some windows api change or lack of success of those Copilot+ PC and 40 Tops npus, has if those meant anything. 2033/2040/2050 AI inference on perscnal computer will not be based on those choice at all. Same goes for what agentic local work agents will look like, nothing to do with Microsoft early copilot type attemps, openCLAW like using multimodal models and many experts/size for different things working together is much more the direction it seem to be going (which will be able to use ASIC for AI if they get popular at home, the community will make it happen quick, can already work on NPU but those ai pc type are not really worth the trouble).

And do not use as an influence any of the pundits that have been wrong at every possible turn since 2023 on the question (the we will not need much compute once the model are trained type), they have no idea what they are talking about, not a single clue.

LukeTbk said:

outside gamers, who think that expensive working local AI agentic is for ?

Click to expand...

LLM-based agents are for people who don't mind their databases deleted once in a while.

LukeTbk said:

local data laws-regulation

Click to expand...

AFAIK private cloud complies with all of those regulations. The only exception is top secret, which is very rare.

LukeTbk said:

People that work on computers is not a niche market(that a giant proportion of the world workforce in the western world that what 33-45% of people ?)

Click to expand...

It is niche outside of potato level enterprise PCs with tiny cases which can't even fit a GPU. Most enterprise customers need Excel, Outlook, Teams. All of which run well on integrated graphics.

DukenukemX said:

Point is AI is rarely if ever used locally

Click to expand...

There is sizable group who use stable diffusion to generate foxgirl boobas. We all know it's a $6 trillion market.

In 10 years, the datacenters will start getting rid of all of those fancy $30K boards. We'll be able to find one in good condition for $2,000, then local models make sense.
Abode will be forced to do freemium like Affinity. Media production niche will surely be disrupted.
The problem is that there's a huge surplus of human labor in media generation. Another problem is lack of novelty and saturation. Nothing new in music since the 90s electronic boom, and that was largely thanks to coupling it with MDMA. UC Berkley did a survey that showed that 60% of gamers aren't willing to pay $70-80 for a game. AI is disrupting a stagnant (currently shrinking) market.

AI is still the best growth story in the last 20 years, but as of now there is no path to consumer adoption.

stuffmonger said:

There is sizable group who use stable diffusion to generate foxgirl boobas. We all know it's a $6 trillion market.

Click to expand...

That's not what most people do with their computers.

In 10 years, the datacenters will start getting rid of all of those fancy $30K boards. We'll be able to find one in good condition for $2,000, then local models make sense.

Click to expand...

This guy got a V100 for $100, which is a 9 year old data center GPU. That's even during a healthy AI market, which we know won't last forever.

View: https://youtu.be/7DAPd5MGodY?si=6LSwDzZDV6zNTZNm

Abode will be forced to do freemium like Affinity. Media production niche will surely be disrupted.
The problem is that there's a huge surplus of human labor in media generation. Another problem is lack of novelty and saturation. Nothing new in music since the 90s electronic boom, and that was largely thanks to coupling it with MDMA. UC Berkley did a survey that showed that 60% of gamers aren't willing to pay $70-80 for a game. AI is disrupting a stagnant (currently shrinking) market.

AI is still the best growth story in the last 20 years, but as of now there is no path to consumer adoption.

Click to expand...

AI disrupting art is already a thing. Once people get access to some good LLM's, then things will get out of control. The reason AI art is hated is because a lot of people make money from it. Just look at DeviantArt and NewGrounds, and you'll see that people just pump out art to try and make money. But, most people who view art won't care. That's not to say an AI artist can't still make money, but it will get ugly as AI artists can pump out more and likely boring art.

DukenukemX said:

This guy got a V100 for $100, which is a 9 year old data center GPU. That's even during a healthy AI market, which we know won't last forever.

Click to expand...

A100 with 80GB still go for over $7,000 on ebay (and those tend to go for much higher than that, specially 32gb version), GPU with enough ram will be used for the pre-filling phase that feed NPUs inference for a long time, H100 will probably have a ~10 years issh lifetime on the very high end use, but used along only to do inference by itself, the $2000 amount of ASIC of 2036 should destroy them on low bits quantizisation (or the equivalent of the days)

stuffmonger said:

LLM-based agents are for people who don't mind their databases deleted once in a while.

Click to expand...

cloud based agents can do to a lot of damage true access-users control and so on, Linux give you a lot of control, Microsoft--nvidia are working hard to make it easier on windows has well:
https://github.com/microsoft/mxc
https://github.com/NVIDIA/OpenShell
https://developer.nvidia.com/blog/b...pcs-with-new-tools-from-microsoft-and-nvidia/

create a close sandbox for your agents that do not know the rest of the system exist, making it hard for them to mess the database backups.

And if you run your session on the cloud on a encrypted way, the agent become local to it (and they will still use NPU even more often)

DukenukemX said:

AI disrupting art is already a thing. Once people get access to some good LLM's, then things will get out of control. The reason AI art is hated is because a lot of people make money from it. Just look at DeviantArt and NewGrounds, and you'll see that people just pump out art to try and make money. But, most people who view art won't care. That's not to say an AI artist can't still make money, but it will get ugly as AI artists can pump out more and likely boring art.

Click to expand...

It's much harder to make good money from selling art directly than from working a job. Most art is bought as a part of a bigger package, e.g. a game.
Here's a job posting that requires Stable Diffusion.
https://www.neowiz.com/en/career/browse-job/043f65b9-9c3e-4063-9a86-d73eccae503e

Artists are butthurt because companies now require them to become technical artists, which means doing boring nerd shit they hate.
There's also pure art that doesn't need to be sold. AI is another tool for creating digital media.

View: https://www.youtube.com/watch?v=RjdeE_TJPko

Something like this is close to impossible to make without video models. It took hundreds of prompts, and edits. It's novel and fun. The guy who made it is more of an artist than someone who draws concept art for some corpo by hand.

LukeTbk said:

create a close sandbox for your agents that do not know the rest of the system exist, making it hard for them to mess the database backups.

Click to expand...

An agent in a sandbox with no agency. That would be great ... unfortunately, I see no reason to do something boring and retarded. I'm in the wrong audience for this, but surely, you can find the right crowd on YouTube.

AMD and Intel arm x86 against the AI gap with ACE, baking matrix-multiply engines & low-precision formats straight into future CPUs​

Hassan Mujtaba
Jun 19, 2026 at 08:05am EDT

ACE, the upcoming set of x86 Extensions defined by both AMD & Intel, has seen the latest spec release, focusing on AI acceleration.

AMD & Intel Focus on AI Acceleration Through Next-Gen x86 Architectures That Are ACE Compliant​

the latest ACE "AI Compute Extensions" specifications have been published by AMD and Intel, which give us an insight into what this new feature for x86 chips has to offer.


https://wccftech.com/amd-intel-arm-...engines-low-precision-ai-formats-future-cpus/