News for hackers. Stuff that matters.
Re:Snapdragon X70 by sitzkrieg on Tuesday March 01, 2022 @07:38PM (#30522069)
call me stupid but every single thing listed in the AI section doesn't require AI to do well?

some graph optimization, i guess. but channel, antenna tuning, network selection? what, isnt this all simple closed form "pick the best number"?

if its just flexing its ability to be used as edge compute as well thats cool too

Re:Snapdragon X70 by kimixa on Tuesday March 01, 2022 @09:05PM (#30522788)
Stuff like locating stuff for beamforming or MIMO has a non-zero cost, and newer standards are getting even more complex trying to drag signal out of noise and error correction and stuff.

The first is exactly the sort of thing where a cheaper "Good Enough Estimate" may be a great solution - which is what a fair bit of "AI" hardware does today. Though I don't know if firing up a separate AI block instead of an integrated to the radio "hardware" solution would be better, but it may be there's some crossover of techniques that can help.

Re:Snapdragon X70 by MPSimmons on Tuesday March 01, 2022 @09:47PM (#30523039)
They use the AI buzzword a WHOLE lot, and don't explain at any point what they mean or how the model (presumably) was trained, or on what data sets, or really anything.

It's just a vacuous statement at this point.

Re:Snapdragon X70 by zibzab on Thursday March 03, 2022 @07:43AM (#30540162)
I am all for being sceptical to corporate PR, but Nvidia and AMD have already proven that AI can be applied with excellent results.

How it's done is a business secret, I would be surprised if they told us much more.

Re:Snapdragon X70 by elil17 on Tuesday March 01, 2022 @09:32PM (#30522965)
Pick the best number is hard when the “number” is something like what direction to beamform in. That’s a complex function of the processors physical surroundings. It would not surprise me if an AI trained on a lot of real world data does better than a hand tuned algorithm in the average case.
Re:Snapdragon X70 by ksec on Wednesday March 02, 2022 @02:19AM (#30524702)
>It would not surprise me if an AI trained on a lot of real world data does better than a hand tuned algorithm in the average case.

I have never read a single comment that appreciate the complexity of 3GPP modem design. So this may be the first one in ten plus years.

Designing a working 4G / 5G modem is not hard. Designing one that is as good as Qualcomm and works in thousands of edges cases in hundreds of different carrier configuration all around the world is god damn hard.

Massive-MIMO on FDD in itself is complicated enough. One could argue we should all switch to TDD, but we dont have that luxury and it is not always a win. Carrier aggregation with mmWave beamforming since Qualcomm is somehow still keen on getting it done.

And this is just on the client side. Most if not all on HN have very little to no appreciation on the complexity of mobile network and 5G. No wonder why sometimes you hear engineers working in telecom equipment industry are rather pissed about Tech industry in general.

Re:Snapdragon X70 by elil17 on Wednesday March 02, 2022 @09:35AM (#30527738)
Absolutely! I worked on 5G in a non-engineering role for a few months and was totally blown away by the amount of complexity in modern modems. Also amazed by how many opportunities for improvement those engineers found/how far away we are from a “perfect” system, especially with Wi-Fi.
Re:Snapdragon X70 by fao_ on Tuesday March 01, 2022 @10:18PM (#30523250)
I would be, given that "AI" hasn't really performed very well on any of the problems we have thrown at it thusfar. Perhaps speech generation is the best fit, but outside of a small smidgen of realms "AI" is not really worth the cost compared to simpler strategies that not only have higher accuracy but a more predictable mode of failure that can be designed around and ameliorated.

https://www.technologyreview.com/2021/07/30/1030329/machine-...

https://www.wsj.com/articles/walmart-shelves-plan-to-have-ro...

https://medium.com/starsky-robotics-blog/the-end-of-starsky-...

Re:Snapdragon X70 by rsashwin on Tuesday March 01, 2022 @07:44PM (#30522129)
That stood out to me as well. I feel it using AI for simple applications like these is a bit unnecessary.
Re:Snapdragon X70 by robbedpeter on Tuesday March 01, 2022 @08:40PM (#30522581)
If it saves significant power by getting a "good enough" answer over the energy needed to get the exact answer, it could be very practical.
Re:Snapdragon X70 by charcircuit on Tuesday March 01, 2022 @08:40PM (#30522584)
Collecting data and doing calculations on it takes time and the environment is often not static.