Rowhammer

Conundrum, Tue Mar 10 2015, 05:50AM

Hi, did anyone see the article suggesting that nearly all laptops could be vulnerable to this attack?

I was more intrigued with the possibility of implementing an AI on a standard x86 laptop using 2GB RAM and a bootable DSL pendrive which would be unprecedented.

see below.

"
Re. RowHammer

Interestingly I suggested quite a while back something along these lines to implement a neural net using Flash memory, and actually have some schematics here for an AI that uses this exact technique to get nearly-quantum level speedup effects using a bootable pendrive that runs DSL and then uses the leakage between the memory cells (has to map out chips and look for correlations but that is doable) to run the NN."

my original post on using inter-cell gaps on large Flash chips to implement an analogue neural net was quite a while back, wonder if anyone has tried it yet?

The same technique should also work on the RaspPi Model B+ as it has the same 1GB DDR3 chip as most RAM and is pretty simple to overclock by changing a single crystal to a programmable oscillator.
This would obviously break the HDMI and composite outputs due to timing issues but for such tasks you would simply drop the clock back down to stock for readout or use a second Pi or Arduino just for readouts.

The CPU on the Pi should run best at a low (ie 7C) temperature and the RAM chip kept at +65C with a feedback loop to
keep it on the edge of instability without it fully crashing.

EDIT: Got a response back but have been asked to keep the exact wording secret until I have working exploit code.
Seems that this approach is actually used on some variants of the Block Erupter to get slightly higher performance with lower core number by using two CPUs to check each other's work.

EDIT 01/09/15:

Have working prototype on old DDR2 based laptop, seems to behave as expected.
the trick is to run it from the cpu cache only with a high speed low density module in position 1 and the 2GB in position 2 on the extensa 5xxx which allows DSL to work without crashing.
then map the 2GB as high speed low refresh for maximum entropy.

Re: Rowhammer
Conundrum, Wed Oct 05 2016, 06:17AM

Hi all.
I am not sure why, but pretty sure this wasn't actually all my fault as the problem on my x520 suggests that the BIOS chip on some laptops incorrectly reads the SPD data (serial presence detect) upon resuming from sleep.
The issue here is that some netbooks and for that matter many commercial laptops even production models are *very* sensitive to voltage spikes over the USB eg from plugging in an external hard drive. The effects can be bizarre to say the least, have experienced subtle corruption of the hard disk that only manifests under very specific repeatable circumstances.
I also observed what appears to be firmware corruption on the drives when looking on a Toshiba c650d (similar DDR3) as key sectors of even a wiped drive show areas that are much slower than normal (50ms rather than 3ms)
If so this could be a new form of BIOS rootkit that can evade all known detection and is worth documenting.

Also relevant, it seems that 28Si is the quantum "magic bean" here and chips from certain geographical locations may be more vulnerable than others. Its possible to test this indirectly by heating up a given batch of modules under rowhammer testing and looking for any showing instability; I did just that a while ago and it is a very effective test. Its worth mentioning that older generation laptops using ecoROHS seem a lot more likely to yield sensitive modules and in fact pc8500s is particularly prone to this.
Overclocking a "good" module on purpose is a very effective way to test it, and map out any suspect areas for high reliability systems eg for servers.
Just be sure to reverse the overclock, it isn't actually possible to break memory this way as it is a quantum effect and thus does not permanently damage anything provided that the chips stay within their Top max.
I got a test stick up to close to twice its rated speed before it finally gave out and threw errors smile