Artificial Intelligence thread

tphuang

General
Staff member
Super Moderator
VIP Professional
Registered Member
I just returned from a business trip, and my development is mainly focused on embedded environments, as well as tinkering with NPU AI applications. The solutions I have been exposed to include
rk3399/rk3568/rk3576/rk3588/ax630/ax650(1-18TOPS int8)
What I am doing in the future will basically involve the targeted application of all domestic AI SOC, including those discussed by the general public and those not discussed
I haven't had much exposure to AI SOC now, purely because I don't have the money to buy too many development boards and carry out development (I am currently working on mass production customization of AI hardware, with an estimated investment of 120000+USD), because I will add NPU/AI functionality to a specific customized Linux solution. It is 10 times more troublesome than on official custom Linux distributions (usually Ubuntu/debian/armbian).


I'm no amateur in AI hardware/software, though my projects primarily use the YOLO series—not LLM-type models. Truth is, current-generation NPUs like those in the RK33/RK35 series SOCs aren’t well-suited for LLMs.

It’s the next-gen RK36 series that’s designed for it.

I respect the information shared by the person in the video, though I wouldn’t consider them an expert either. That said, their insights are particularly intriguing.

I’ve engaged with acknowledged heavyweights in China’s internet circle—many now diving into AI. Technically, I’m confident we have a far clearer grasp of the underlying tech than these industry leaders.

But commercially speaking, early-stage innovation ventures’ survival—fundamentally reliant on a CEO’s fundraising prowess—means individuals like those in the video prove indispensable in the commercialization process.
Yes I have tested with pretty much every board you mentioned in there. And rk3688 won’t be ready until end of 2026 based on my conversation with rockchip guys a month ago. Axera is actually proposing something with higher tops for int4 and they have in fact better software support. They told me their next chip will do 72 tops. But okay, even with 18 tops, I am not fully utilizing the NPU because the memory they put on those boards are too small. So what’s the point?

There is also horizon but there is no LLM support for them.

A huge % of people in Shenzhen think the current SoC can’t run LLMs on edge but guess what, a lot of people in China are trying it out. Do you know how I know this? Axera told me. My former partner factory in Indonesia said this too.

and I am pretty sure you can get it working reasonably. Whether people actually can sell the product is a different story.
 
Top