Hammered...

Harking back to my post of the night before last, I want to mention a phenomenon I was hitherto unaware of until Joe sent me a cryptic text about Nvidea GPU [Graphics Processor Unit] chips now being vulnerable to RowHammer bit-flip attacks. As is my wont, I decided to at least attempt to plug this gap in my knowledge by digging around the InterWeb™for information. I soon lit upon a couple of academic papers from Carnegie Mellon University and The University of Toronto: the first a retrospective paper concerning researches carried out by the authors, published as early as 2014, and the second published as recently as July this year on the phenomenon having been proven to be usable in attacks on Nvidea GPUs, the chips used widely throughout the computing industry for gaming, encryption, blockchain farming and AI.

The practical upshot from these researches is that there is a known fundamental vulnerability to 'bad actor' attacks, using a particular characteristic of modern computer dynamic memory to either disrupt and directly disable a computer or embedded system, or to, more insidiously, inject malicious code into memory for execution as if it belonged to the software running in that system's memory. This can be achieved through the use of code as trivial as a JavaScript running in a web page. The vulnerability, ironically, is as a result of the ingenuity of hardware developers managing to squeeze more and more components onto silicon chips to enable them to work ever faster and more energy-efficiently. At the kind of scale now possible and routine, memory chips exhibit what could be termed 'self-interference'. Dynamic memory has to constantly 'refreshed' every few milliseconds in order to hold its data. If a particular 'row' of memory is refreshed often enough, it can interfere with adjacent rows, causing them to change state: 1 to 0, 0 to 1, corrupting their data in turn.

This phenomenon happens in quite normal operation and its effects can be partly mitigated through various statistical software/firmware techniques, but it turns out that this behaviour can also be instigated remotely and used maliciously against any suitable targeted system. Until now, this was thought only to be a problem  particularly affecting embedded dynamic memory in CPUs - [Central Processor Unit] the main processor of any computer or mobile device or whatever; but the paper that emerged from The University of Toronto this year demonstrates that similar attacks on GPUs are in fact a practical proposition and it turns out that the most widely used GPU chips - Nvidea's - are in fact vulnerable to such attacks via this route. This should be a cause for great concern, as these devices are almost ubiquitous in not just supercomputers - the new one at Bristol alone employs thousands of them - online encryption and large scale data-wrangling such as in AI, but also in millions of networked devices such as home computers and games consoles. The potential for mass manipulation of data to nefarious ends is practically limitless, and the consequences dire, so wholly reliant we now are on these technologies in every aspect of our lives. Perhaps we might heed the work I referred to in a post earlier this year to prepare for the inevitable...