Ai chips and chip makers are trolling the consumers 🤣🤣🤣

I bought this mini pc with the latest ai chip the Ryzen ai 9 470hx and it’s a Gimmick!!! Do the chip makers hate us or are they out of touch with reality. Apple is the only company that seems to get it. Or maybe they do get it but just want everyone to spend 7k for a 5090 🤣🤣. We just need faster bandwidth speeds. No fancy useless NPU. No reason why in 2026 with AI being the focus, nobody figured out the bandwidth speeds bottleneck. The biggest bottleneck for everything.

The only people who are out of touch are those who spend hundreds, or thousands of dollars on AI hardware without doing research. There are dozens of reviews on the AI performance of the ai 9 470hx. There should be no surprise on how it performs.

If this is the latest and greatest I can only imagine the rest 🫣🫣🫣

We know how to increase memory bandwidth but client AI isn't a big enough market to be worth fundamentally changing the architecture from normal DDR5. Apple did it long before the AI boom started so it's mostly just a coincidence that their architecture is also good for AI use cases.

Apple also includes NPUs are their SoCs if you weren't aware.
Faster BW isn't in the cards for any mainstream AMD or Intel APU, because memory channels are expensive.

And NPUs are FAR more efficient at the tensor compute required for Microsoft's vision of "local AI".

AMD does have a higher BW part and it's the Ryzen AI Max+ 395, it's just useless and expensive for most people whom just buy simple laptops.

You'd be surprised just how much Zen3 parts AMD still sells in 2026 simply because they are cheap to manufacture and perform "good enough" for most laptops.

I mean maybe if they increase it, they would attract more people. And it would become big enough. Any PC chip maker that increase would own the market. That’s the biggest gap in ai and the biggest bottleneck. All the chips even the older ones can run a llm efficiently. Biggest issue seems to be bandwidth speed. Which is absolutely insane in 2026 🤣🤣🤣

I get it…my asus tuf with a 7735hs and a 7700s perform better than all of the ai chips they have now…and I got it open box for $840…got 2 16gb ddr5 for 180 on facebook marketplace a month ago…now that same laptop cost more than when I bought it 3 years later 😭😭😭. Only issue is the speed once a model start offloading 2 much because of the bandwidth speed between data transfers between gpu and cpu…all the computing and complex maths get done fast…just the transfer to start generating tokens.

Well yea man, you’re comparing a power hungry discrete GPU, over a tiny block in a laptop designed to consume 2-3w of power to deliver its local AI inference.

You got your dual channel DDR5 at 5.8Gbps or whatever, against something running 18Gbps GDDR6, you’re not going to get magic.

I mean I expected Atleast 10-12 tk/s and I didn’t expect the models to become slower the larger they get no matter how much ram you have. Trial and error, now I know. Might as well build a pc or get a laptop with a dgpu if you can get one under 1k. Mini pc useless unless it’s a Mac with Atleast 24gb ram…but to honest qwen3.8 9b q4 is amazing…and if you know how to to make your own tools and skills it can do a lot.

# 6. Serious and intelligent discussion; No opinion posts.

We allow discussion posts on relevant topics to hardware; there should however be at least background information and relevancy to demonstrate such a discussion is worthwhile. Low effort posts such as "What is \_\_\_\_\_\_ and what does it do?" is not acceptable.

Opinion posts are generally not allowed, unless there are demonstrably strong relevancy to hardware as well as substantial evidence provided.

The vast majority of AI users just want to use an app that connects to an AI cloud service, they don't care about local models.

Even in a normal hardware market it wouldn't make sense to use something like HBM for the tiny market of local AI models, but in 2026 with the AI cloud companies having completely destroyed the memory market, it would be completely infeasible to do what you want at a price anyone would be willing to pay.