AMD Zen 6 desktop CPUs may ditch integrated graphics for a built-in NPU

How large is this NPU going to be? Because removing the iGPU frees up a considerable amount of silicon (1WGP+command front-end with geometry processor+ACE+render backends at memory controllers). The media engines could be wired to CPU-side and simply directly access memory controllers in IOD, but generally, are also part of the iGPU. That's a massive hole in the die now.

I think something was lost in translation. It's more likely that AMD will use a larger 4-8CU iGPU that has specialized NPU functions, like a hybrid iGPU+NPU. NPU in my laptop sits unused 99% of the time. Graphics engines could be gutted, not unlike CDNA, I guess ...
 
Last edited:
@passwordistaco's ancestors:

1877: "telephones in homes are not useful."
1884: "electricity in our cities is not useful"
1908: "horseless carriages are not useful"
1975: "personal computers are not useful".
1995: "The Internet is a fad that will soon pass."

A rather silly analogy, especially considering that it's straw man argument. But even assuming that it was said like you claim, a personal computer bought in 1975 was indeed not useful for most people, and was a waste of money for them. That's the point. Nobody is saying that AI isn't useful, but do you think that an integrated NPU next year would actually end up useful for anyone? I'd say that it's not even clear if an NPU will ever be useful, considering that internet-based AI will always be significantly more powerful.
 
Uff not good for business PC's. We have several computers with 9900X, with using integrated graphics.
 
A rather silly analogy, especially considering that it's straw man argument.
I don't believe you know what a straw man argument is: his claim was repeated verbatim, merely translated into different eras.

...Nobody is saying that AI isn't useful...
You haven't been on these boards long to say that.

...but do you think that an integrated NPU next year would actually end up useful for anyone?
Was that a joke? Already more than 1.5 billion people have used some form of AI tool directly, and at least 350 million use them on a daily basis.

That's today. Analysts estimate that number to double within 18 months -- and a chip released in 2027 generally targets at least a three-year window after that.

...I'd say that it's not even clear if an NPU will ever be useful, considering that internet-based AI will always be significantly more powerful.
Funny, your remark mirrors almost exactly that of an IBM CEO who claimed PCs would never be useful, because mainframes were so much more powerful ... and indeed, when IBM released their original PC, they famously underestimated demand by orders of magnitude. That underestimate is why they outsourced the PCs software to a tiny company named Microsoft, in fact.

Oops.
 
I build price conscious units for business and schools. Integrated graphics help keep the costs controllable. Adding the current overpriced GPU just increaces costs and possible troubles.
This is a bad move if this is really what AMD is going to do.
 
The iGPU on a desktop Ryzen has never once been used for gaming. Its entire existence is for the two moments in a PC's life: initial build POST and that one time your GPU died and you needed to limp through ordering a replacement. Removing it to add an NPU that nobody asked for is a genuinely strange trade-off.
 
I actually dislike integrated GPU in a desktop system where I'm using a discrete GPU. As a Linux user, I prefer a single discrete GPU, and I actually search and buy a desktop CPU that doesn't have integrated graphics.

But for laptops, I prefer integrated graphics.

Integrated graphics have been a lifesaver when the discrete GPU fails. Over the years I have had at least 3 discrete GPUS fail. Switching to integrated immediately determines the problem.
 
@passwordistaco's ancestors:

1877: "telephones in homes are not useful."
1884: "electricity in our cities is not useful"
1908: "horseless carriages are not useful"
1975: "personal computers are not useful".
1995: "The Internet is a fad that will soon pass."

Intgrated NPUS will probably find a niche as a private porn experience provider.
 
A Good Enough GPU like 780m in Ryzen 7 8700G Helped me in Choosing AM5 Socket Mother Board Instead of AM4 MoBo As I Could Avoid Buying a GPU for the Initial Build Saving me $150. Given The Absurd Increase in Ram and SSD Prices Having an APU with Decent Good Enough GPU ( 8-12 CU or More ) Helps in Building Affordable PC.
 
I'll admit I haven't a clue what an NPU would offer me. I see adverts for AI agents and the like but have no clue what they're actually offering. It all seems pretty much like snake oil. The benefits of an integrated GPU, on the other hand, seem obvious.
Clearly - Really Don't Get the Hype Of NPUs. Maybe they are only useful for Super Productive People. Other than that I don't see Normal Folks making Use of NPU that much. Most of Them Will Just use Free AI or avail Basic Subscription of AI.
 
Was that a joke? Already more than 1.5 billion people have used some form of AI tool directly, and at least 350 million use them on a daily basis.
I know what a straw man argument is. This is an example. I said that AI is useful but I doubt an NPU currently is. You obviously don't have any example of an NPU being of actual use, so you answered as if I said something about AI not being useful.

That's what people without a good argument do. My guess is that you don't even understand the limitations of an NPU vs. cloud based AI, and that you're in generally ill-educated on the subject of AI.
 
I know what a straw man argument is. I said that AI is useful but I doubt an NPU currently is
NPUs run AI models. Your statement is like claiming gas-powered lawn motors are useful, but gasoline is not.

... My guess is that you don't even understand the limitations of an NPU vs. cloud based AI, and that you're in generally ill-educated on the subject of AI.
I've been involved with predictive ML since the days of genetic algorithms and simulated annealing, more than 30 years ago. You? My guess is you fail to understand orders-of-magnitude compute difference between training a model vs. using it for inference. No one's saying that a budget NPU is going to host multi-trillion parameter LLMs ... but that's a small fraction of the potential market.
 
Last edited:
Back