I've looked at Visual 6502 and it's way beyond me. I've even looked at the scans where it shows how they severed the connection to disable 6502 decimal mode on the NES.
Getting a logic-gate-level netlist from a GDS is trivial with industry standard chip design tools. Circuit designers do this every day. The hardest part will be reverse engineering the functionality.
I wish I had more time and I'd throw Calibre at it.
Indeed, this is a problem for an energetic student who has free access to every tool in the Cadence, Synopsys, and Mentor portfolios: take the GDS-II layout, run it through extraction, generate a netlist, convert to higher-level blocks ("netlist-to-gate recovery"), translate to RTL HDL, then attempt to work out "what it does."
This is just a job posting in disguise for Jane Street: solve the puzzle, get a call from one of their recruiters.
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.
> He did that by connecting oscyloscops to all chip's outputs and started with applying random current on inputs
This is absolutely not how reverse engineering a digital logic ASIC works.
Either the story got embellished through retellings, or this person was a fantasist.
There are people who hack on GPUs but it’s done at the software level.
I did get a kick out of imagining a scene where someone is trying to connect an oscilloscope to a circuit board to reverse engineer the CPU opcodes. That’s like the CSI: Miami version of what this would look like.
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.
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)
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.
Also a few free-but-cut-down versions like 5 (which still recognizes MS-DOS executables). I'm not going to recommend downloading it from any of the locations mentioned here but it's out there: https://reverseengineering.stackexchange.com/questions/19179...
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.
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.
Yes. There is exactly that, and we call it an "extraction" tool. It takes a GDS (text representation of shapes in the physical layout), and gives you back a "netlist" (text representation of components and connections in a circuit schematic).
Circuit designers use these tools basically daily for two reasons - the first is Layout Versus Schematic. We want to make sure that the physical layout matches the schematic, so the tool turns the layout GDS into a netlist and compares that to the netlist created from the schematic (basically a diff, but more complicated). The second is so we can run simulations that take into account the "parasitic" resistances and capacitances of the wires and metal shapes in the physical layout. It's basically the same procedure as LVS with an extra step that analyzes the metal shapes to determine said R's and C's.
The tools can definitely recognize static logic gates purely from the topology (nothing special in the GDS needed). I don't know that I've tested it for anything more complicated, but as you've pointed out, if the hierarchy is in the GDS it's trivial to recover
In fact if you have the sky130 libraries installed (it's where I'm working) you can get a spice netlist out of it in about 10 seconds - really they should have flattened it
Practically No, the stack-up of metal layers often hides the gate structures underneath, and the billions of process cells may not all be the same.
Theoretically Yes, as an ion-beam-mill and electron-microscope combination machine can slice up semiconductors layer-by-layer. Given these machines can often also give precise x-ray analysis material data, the exact makeup of the chip can be extracted by competitors given enough time. =3
I was thinking of the kind of high-energy photon sources that sprawl across a large laboratory campus behind rings of security fences.
These things can definitely erode the targets. Tomography experiments have to think carefully to optimize the set exposure angles used during a session, because the target degrades more with each shot.
From where do I know the name Dragon Sector and q3k? You aren't the ones who hacked the train DRM, are you? Or maybe active in the demo scene? Or maybe I'm just confusing you with TRSi?
I'm a chip designer, back in the early 90s I visited a small silicon house who were doing some work for us, their boardroom table had been taken over by a giant sized photomicrograph of IBM's VGA chip and they had a bunch of summer hire grad students tracing wiring and matching standard cells (easier back then because we only worked in 2 layers of metal).
They weren't trying to steal IBM's design, they already had one of their own, what they wanted to know was "is there any as yet unannounced functionality in the IBM VGA chip?" they didn't want to be caught out when some unknown registers popped up. Of course it turned out there was as yet unannounced functionality - but not by IBM's design, more because of the orthogonality of the design - MODE-X was discovered and became the basis for DOOM's speed, anyone who didn't support it lost out
They did give the chips to China, but they didn’t work. Turns out they’d first taken them to the basement where someone cut them into pieces before sending them to China.
I thought the article was going to be about how people scan chips
Rapid Chip Reverse Engineering Using Laser, Focused ion beams, and Scanning electron microscope https://academic.oup.com/mam/article/30/Supplement_1/ozae044...
FIBs are also used to test modifications before doing a respin. I'm still in awe that matter can be manipulated so precisely
Man, these guys always have insane puzzles. What the heck.
Hasn’t Ken Shirriff been doing this for quite a while? (:->
I'm too busy trying to reverse-engineer the Intel 8087 floating-point chip, although it would be interesting to look at this puzzle chip.
I've looked at Visual 6502 and it's way beyond me. I've even looked at the scans where it shows how they severed the connection to disable 6502 decimal mode on the NES.
To be fair, the 6502 is very dense, and very hand optimized.
Modern systems tend to use "standard cells" for logic which is a lot more digestible.
Getting a logic-gate-level netlist from a GDS is trivial with industry standard chip design tools. Circuit designers do this every day. The hardest part will be reverse engineering the functionality.
I wish I had more time and I'd throw Calibre at it.
Indeed, this is a problem for an energetic student who has free access to every tool in the Cadence, Synopsys, and Mentor portfolios: take the GDS-II layout, run it through extraction, generate a netlist, convert to higher-level blocks ("netlist-to-gate recovery"), translate to RTL HDL, then attempt to work out "what it does."
This is just a job posting in disguise for Jane Street: solve the puzzle, get a call from one of their recruiters.
It's even easier in this case because they've included the original verilog source ....
No, that's for the warmup. The real puzzle only has a .gds file available.
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.
> He did that by connecting oscyloscops to all chip's outputs and started with applying random current on inputs
This is absolutely not how reverse engineering a digital logic ASIC works.
Either the story got embellished through retellings, or this person was a fantasist.
There are people who hack on GPUs but it’s done at the software level.
I did get a kick out of imagining a scene where someone is trying to connect an oscilloscope to a circuit board to reverse engineer the CPU opcodes. That’s like the CSI: Miami version of what this would look like.
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.
Small pedantic nit: Open-sourcing is not the same as reverse-engineering.
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)
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.
Also a few free-but-cut-down versions like 5 (which still recognizes MS-DOS executables). I'm not going to recommend downloading it from any of the locations mentioned here but it's out there: https://reverseengineering.stackexchange.com/questions/19179...
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.
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.
“oscyloscops” is a way better spelling I gotta say.
What prerequisites I need to solve this puzzle? I don't have much knowledge in electrical engineering.
Is there something like an Extract-SPICE tool that takes a circuit and gives you back a text rendering of it ?
Yes. There is exactly that, and we call it an "extraction" tool. It takes a GDS (text representation of shapes in the physical layout), and gives you back a "netlist" (text representation of components and connections in a circuit schematic).
Circuit designers use these tools basically daily for two reasons - the first is Layout Versus Schematic. We want to make sure that the physical layout matches the schematic, so the tool turns the layout GDS into a netlist and compares that to the netlist created from the schematic (basically a diff, but more complicated). The second is so we can run simulations that take into account the "parasitic" resistances and capacitances of the wires and metal shapes in the physical layout. It's basically the same procedure as LVS with an extra step that analyzes the metal shapes to determine said R's and C's.
yes, but at the transistor level .... it has no idea at that level how those transistors are arranged into gates directly
The tools can definitely recognize static logic gates purely from the topology (nothing special in the GDS needed). I don't know that I've tested it for anything more complicated, but as you've pointed out, if the hierarchy is in the GDS it's trivial to recover
I take it back, it's a sky130 .gds the gates are explicitly in there
In fact if you have the sky130 libraries installed (it's where I'm working) you can get a spice netlist out of it in about 10 seconds - really they should have flattened it
Practically No, the stack-up of metal layers often hides the gate structures underneath, and the billions of process cells may not all be the same.
Theoretically Yes, as an ion-beam-mill and electron-microscope combination machine can slice up semiconductors layer-by-layer. Given these machines can often also give precise x-ray analysis material data, the exact makeup of the chip can be extracted by competitors given enough time. =3
Now you're making me imagine some kind of 3D-scanning, confocal x-ray fluorescent spectroscope.
Or maybe some kind of hybrid of x-ray microtomography and spectroscopic analysis all in one.
But, maybe the energies involved would be about the same destructive power as some microtome slicing technique...
We already know that X-rays don't destroy chips.
I was thinking of the kind of high-energy photon sources that sprawl across a large laboratory campus behind rings of security fences.
These things can definitely erode the targets. Tomography experiments have to think carefully to optimize the set exposure angles used during a session, because the target degrades more with each shot.
looks like they didn't post any blog post about 2nd NN challenge (https://huggingface.co/spaces/jane-street/droppedaneuralnet)
I was waiting for some writeup about permutation decyphering
30 minutes with /goal for the solution from Sol w/ high.
In a simplified scenario (not too far from this)? Yeah, we've done that in CTFs almost a decade ago.
https://blog.dragonsector.pl/2017/10/?m=1
From where do I know the name Dragon Sector and q3k? You aren't the ones who hacked the train DRM, are you? Or maybe active in the demo scene? Or maybe I'm just confusing you with TRSi?
Maybe. :)
That sounds like a fun challenge. Feels a lot more tractable than the neural net one.
people who can do this stuff are super-smartypants
but reminds me how we're going to find out on an industrial level when the Saudis give China some nvidia chips they were grifted
they've cloned lots of chips before but nothing that advanced
I'm a chip designer, back in the early 90s I visited a small silicon house who were doing some work for us, their boardroom table had been taken over by a giant sized photomicrograph of IBM's VGA chip and they had a bunch of summer hire grad students tracing wiring and matching standard cells (easier back then because we only worked in 2 layers of metal).
They weren't trying to steal IBM's design, they already had one of their own, what they wanted to know was "is there any as yet unannounced functionality in the IBM VGA chip?" they didn't want to be caught out when some unknown registers popped up. Of course it turned out there was as yet unannounced functionality - but not by IBM's design, more because of the orthogonality of the design - MODE-X was discovered and became the basis for DOOM's speed, anyone who didn't support it lost out
MODE-X launched 1,000 ships too with the demoscene.
They did give the chips to China, but they didn’t work. Turns out they’d first taken them to the basement where someone cut them into pieces before sending them to China.
You can do this. If you commit the whole next month to it you'll make quite some progress. But you won't.
China has no shortage of Nvidia chips. It costs nothing (relatively) for someone to just buy a 5090 off the shelf and send it there.
Is this the chip they used to steal money from the Indian stock market until they got banned from India?