[xvfLJfx9]

You can't lol. How do you wanna restore a blown fuse on nanometer level INSIDE the GPU die. Its simply not possible.

By the way, AMD also uses fuse blowing if you e.g. overclock some of their CPUs to mark them as warranty voided. They give you a warning in the BIOS and if you resume a fuse inside the CPU gets blown that will permanently indicate that the CPU has been used for overclocking (and thus remove the warranty)

[bri3d]

> You can't lol. How do you wanna restore a blown fuse on nanometer level INSIDE the GPU die. Its simply not possible.

I wouldn't dismiss this so aggressively.

Frequently (more frequently than not), efuses are simply used as configuration fields checked by firmware. If that firmware can be modified, the value of the efuse can be ignored. It's substantially easier to implement a fused feature as a bit in a big bitfield of "chicken bits" in one-time programmable memory than to try to physically fuse off an entire power or clock domain, which would border on physically irreversible (this is done sometimes, but only where strictly necessary and not often).

[Kirby64]

NVidia is smarter than this - they sign all their firmware, so you can't just modify the firmware and bypass this. No signed firmware means no functioning card. The famous example of this was the 'reduced hash functionality' RTX 3060 cards that accidentally had the 'reduced hash' feature disabled in a signed copy of firmware that Nvidia released. If they hadn't accidently released this firmware, the reduced hash stuff would have worked forever.

[bri3d]

I am indeed well aware of how firmware validation works. Finding a vulnerability in firmware validation is however much more likely than reversing OTP for almost all varieties of OTP, even NVidia’s firmware validation which is generally regarded as pretty strong.

[nixosbestos]

This reminds me of when Motorola introduced the first carrier-enforced, signature lock on Android with no "OEM Unlock" option and a bunch of ... questionably informed... people insisted it was by passable. (To my knowledge, it has not been bypassed, and jesus that was 15 years ago probably)

Granted, it's Nvidia, and they've been featured in devices that were notoriously hackable, but also, it's not 2018 anymore.

Needless to say, people should understand when they buy an Nvidia card, they should fully expect to use Nvidia firmware, with whatever that entails.

EDIT: I'd really like to remember the name of this device. It was the same era as Blackberry releasing... the Storm? Some resistitive-touch device with a physically clickable screen. Motorola Storm? I really wish I could recall. (sub-edit: I think the Storm was the Blackberry device. So something else...)

[bri3d]

Maybe the Fire? It looks like a Storm. I don't think they made any resistive Android devices though, even the original Droid was capacitive.

I think there might have been one or two OMX devices in this era with locked bootloaders that weren't bypassed due to a lack of research, but I actually find this example a bit amusing: early Qualcomm Motorola Android phones were touted as "unhackable" due to their use of fuses (Qualcomm even went on a marketing pitch calling them "Q-Fuses"), but were extremely quickly unlocked using trivial TrustZone supervisor vulnerabilities (iirc, there was an SMC that literally had a write-what-where primitive in it).

[wmf]

This is true but that's why the firmware is signed so you can't patch it.

[guerrilla]

> How do you wanna restore a blown fuse on nanometer level INSIDE the GPU die. Its simply not possible.

Bullshit. There will be hackers in the future who can do it in their garage. Just... not anytime soon.

> By the way, AMD also uses fuse blowing if you e.g. overclock some of their CPUs to mark them as warranty voided. They give you a warning in the BIOS and if you resume a fuse inside the CPU gets blown that will permanently indicate that the CPU has been used for overclocking (and thus remove the warranty)

Emphasis on "some." You can buy plenty of CPUs from them made for overclocking.

[rasz]

Not anytime soon as in 50? 100 years? 4090 is on 5nm node, how about you first demonstrate its possible on something 100x bigger? For example Intel started locking CPU multipliers in Pentium manufactured in 0.8 μm (800 nm), unlock that fuse :)

[Kirby64]

> Bullshit. There will be hackers in the future who can do it in their garage. Just... not anytime soon.

I'll take your bet on this. Silicon designers aren't unaware of this potential vulnerability, and if you want to prevent eFuses from being un-blown, you can design for that. I would place money on there not being any commercially viable way to restore an eFuse in a 4090 die at any point in the future. You can probably do it, but it would require millions of dollars in FIB and SEM equipment and likely would destroy the chip for any useful purpose.

Usually the only useful reason to attempt to recover/read/unblown fuses is to read out private keys built into chips.

[guerrilla]

> You can probably do it, but it would require millions of dollars in FIB and SEM equipment and likely would destroy the chip for any useful purpose.

The price tag and size of these things are what I'm talking about. SOME day it will get much cheaper and smaller. A 4090 will be useless at that point, but I still play with 8086s and vacuum tubes, so...

No point in betting though. We'll both be dead by then.

[Kirby64]

I’d still take the bet. FIB and SEM stuff is highly specialized and miniaturizing it to be obtainable by a garage user seems unlikely even in the distant future. Either way, you still couldn’t take a 4090 and make it functional as a 6000 series. You’d destroy it in the process if it was even possible at all.

[guerrilla]

Nah, it's going to become democratized. It's already started.

There's been a few different stories like this lately:

https://interestingengineering.com/videos/guy-builds-integra...

People said the same thing you're saying now about computers. You're just being silly and forgetting history.

[Kirby64]

Making a hobby etching thing that isn’t anywhere close to the state of the art is cool, but not exactly anywhere close to what is needed to look at a modern chip.

You’re trivializing the challenge of modifying something that is on the order of 50nm wide and specifically designed to not be able to be tampered with.