At my uni, 15 years ago, one postdoc reverse engineered NVIDIA chip and wrote more performant compiler. He did that by connecting oscyloscops to all chip's outputs and started with applying random current on inputs. Using ML and his genius he rediscoverd all opcodes including a few hidden ones. Eventually he got hired by some company that was doing a lot of GPU on supercomputers.

▲

inigyou 2 hours ago | parent | next [-]

You can't do that by applying random inputs to any single-chip GPU - it has far too much state. I can see that perhaps it worked on some of the early multi-chip cards - where one chip was a texture sampler, and so on.

You'll have more luck reverse engineering the software driver first. They're not hidden, you can just open the driver files in Ghidra, the almost-universal tool for open-sourcing proprietary code. Hidden opcodes can be discovered first by just trying all the opcodes you couldn't discover any other way. You only need to go to the physical level if they're really hidden.

▲

sigbottle an hour ago | parent [-]

To be fair, Ghidra was released in 2019 and in general knowledge was still hard to find even back in 2010 I feel (well, compared to 2026 in the age of AI)

	▲	inigyou an hour ago \| parent [-]
		Before that there was, and still is, IDA Pro. Works largely the same but costs a lot, on the order of $1000/seat/year. Useless for hobbyists unless pirated, but reasonable if it's your job. Probably had academic discounts.

▲

tverbeure an hour ago | parent | prev | next [-]

There is absolutely no way that happened. 15 years ago, we're talking Fermi class GPUs and chips with hundreds of millions of bits of on-chip state and much more if you include the DRAM.

You can't tease out the right information by applying random inputs. Which input would you even use? The PCIe interface? You'd first "randomly" need to get past its complex training sequences...

Your postdoc probably wrote micro-benchmarks of some sort. That is a common technique.

	▲	kjs3 39 minutes ago \| parent [-]
		And you don't use an o-scope in anycase, since you'd need...what...a thousand of them to watch all the signals. You'd use a logic analyzer. I think I read somewhere that those older nvidia chips had something like 2000 BGA balls, and Tektronix does make an LA that can scale to 2000-something channels (TLA7000), for a modest US$500k or so. Then you gotta figure how to mount the thing to attach the probes. So...agreed...far more likely there was a software solution of some kind if this happened.

▲

pixelatedindex 31 minutes ago | parent | prev [-]

“oscyloscops” is a way better spelling I gotta say.