HN Explorer

How reverse engineering Silicon With AI?

4 points by ahmetak ↗ HN
Hello,

I am just wondering why not we are building new AI/ML to completely build silicone scheme from images of it ? I mean we can ptychography of any chipset. And there is many Open IP so AI can interpret the circuits.

As you know producing the chip not to much, we can accept to pay the engineering and advertising of them however monopoly of GPU/CPU is not acceptable.

Lets talk and create road map or trash this idea.

5 comments

[ 3.2 ms ] story [ 21.6 ms ] thread
FrenchDevRemote
Most AIs doing image recognition/generation use a very low resolution.

Training models on higher resolution would become exponentially more expensive, while still not being that likely to succeed.

And even if you had the schematics, a foundry would still cost millions to setup, + you wouldn't even know if they were any good before trying them out, it would probably be an unpredictable black box.

ahmetak
Actually thats why we need an AI to generate Verilog / VHDL from images of tomography so it can be testable part by part or whole on a FPGA or simulation.

Many part of the chip is actually replication of same compenent or library like DRAM, ROM, Arithmetic operations so always there should be a pattern. Scanning should also work from low resolution to high resolution part by part to find pattern.

And also foundry is not as much as you think, most of the cost is R&D not the production.

https://www.granitefirm.com/blog/us/2023/04/29/cost-of-chip-... https://www.tomshardware.com/news/tsmc-reportedly-adds-advan...

"60-ish A100/H100 GPUs per wafer" and 7nm Waffer cost is $10k

vineyardlabs
Did you actually read the article you linked? 10k per wafer is the cost in-production, at scale. The tapeout process for a 7nm chip is 10s of millions of dollars before you even get to paying per wafer.
ahmetak
https://www.linkedin.com/pulse/talk-chip-design-tape-out-ver...

https://www.reddit.com/r/ECE/comments/7hjlpg/for_anyone_curi...

You are right for production, I was meaning for verification. MPW is around about 20-30k for 7nm.

In full production even with tapeout cost which is $15M for 7nm chipset cost $1500 per die after sell 10k. Best price for A100 is around $6k in monopoly now.

Even before production, after interpreting images, AI can also collect IP cores and produce new ones in the next steps.