Is Moore's Law dead? | Chris Lattner and Lex Fridman

Please support this podcast by checking out our sponsors:

PODCAST INFO:

CONNECT:
- Subscribe to this YouTube channel

[marcv]

I think the steepness of the curve Jim Keller is trying to climb will only continue to get steeper. I seriously doubt we reach the asymptote of 1000X from here. Maybe in principle, but not in a way that ever becomes practical or cost effective.

[PATRICKJLM]

Of course we have reached the limit in transistors' size. The problem with smaller than 7nm transistors is the quantum tunneling - some electrons can jump through the closed gate, messing the binary code. And the smaller you go, the more electrons are "jumping" ... and as you can imagine, they "appear" all over the place on the other side...
Companies are trying to fix those errors in the binary code with software, but as it looks, they don't work well, yet.

[utubekullanicisi]

Meh, of course developers will increasingly write code that takes advantage of both many-core processors and heterogeneous compute (silicon blocks dedicated to doing just one task i.e ML accelerators, image signal processors, etc.), but the reason single threaded perf. hasn't improved over the last few years is because Intel has stagnated in terms of their process node + other strategical mistakes, and AMD was playing catch up, so Intel didn't feel much incentive to do anything more than just add more cores and increase the frequency. Now not only has AMD become very competitive, but we're also seeing increasing competition from the ARM ecosystem from the likes of Apple and Qualcomm (the former which has increased its performance by 300% over the last 5 years) in laptops and soon desktops to compete with x86, especially in integrated (by that I mean whatever the opposite of 'DIY' is), more compact and/or portable systems.

[rogerc]

110volt chip (with low amps) would be more powerful than 12volt chip (with high amps).

[dapdapstuff]

He doesn't know what he's talking about