Hacker News Reader

Comment by chetangole | Hacker News Reader

> The SU7 will have autonomous driving capabilities, too, thanks to its Xiaomi Pilot platform powered by up to two NVIDIA Drive Orin processors

Interestingly, NVIDIA is making good moves in all the areas related to AI.

View on HN

I wonder how many generations before those CPUs get replaced by a Chinese native competitor. I'm guessing the Qualcomm infotaintment processors will be replaced soon.

High end semiconductors is one area where China has a lot of ground to catch up on. They can make them fast, or cheap, but can't make them fast and cheap yet. They have a similar problem with jet turbines, and most of it is related to materials engineering.

But unlike cold war they have market based economy ... so I think that the west will be unpleasantly surprised how little we will be able to slow them down. Also - programmers have started to learn how to squeeze performance again out of the little hot rocks.

It has worked for a few decades now. While almost anything else is easy to reverse engineer, the trial and error used to come up with advanced semiconductors and jet turbine engines is really hard to discern by looking at end products, and the secrets are heavily locked down in Europe and the states (even Taiwan, Japan, Korea have to import materials and equipment from a very few locked down sources for their own fabs). China will eventually get there, but they are 10s of years and billions of dollars away from it. Because of the scale of the work needed, it is also unlikely to come from private industry, and with public research a lot of that money is going to be siphoned off to corruption.

Cars are an interesting example of where it might not super matter that the chips lag. Even a lot.

If you are dumping kW of power into electric motors and climate control, having an inefficient hot not-great 200W computer or two running in car entertainment might not be super noticeable.

The downsides of needing bigger VRMs to power the cores, and needing more cooling are both real costs, that diminsh the effectiveness of trying to save money on cores. But it is interesting to me that this feels like the one segment where people might not notice 100Ws of egregiously inefficient computing, because it's such a small fraction of the total power budget.

> similar problem with jet turbines, and most of it is related to materials engineering.

I think better modelling techniques and far cheaper and better CNC machining will more than make up for subpar materials.

After all, the blades inside a jet engine already have a melting point lower than that of the internal gas temperatures, and there is no limit to film cooling, as long as you can model it precisely enough.

If I had a few years of my life to spare, I wanted to make a demo gas turbine generator with all the moving parts made of chocolate to demonstrate that effect.

They don't need to fab it right now. Xiaomi isn't banned from using TSMC. So they can design it and get it built by TSMC/Samsung. In long run yeah they'll need independence on fab also. But for this use case 10nm should be good enough for now, which SMIC can make quiet competitively.

There is a lot happening in this area atm - here is a description of Nio’s homebrewed chip:

https://www.notebookcheck.net/5nm-NIO-Shenji-autonomous-driv...

Whole Chinese tech companies are adopting "Nobody got fired for using NVIDIA/QCOM" mindset. Gina is trying her hard to undo that tho.

Reply on news.ycombinator.com