Modern CPUs contain a tremendous amount of state that they need to track. Hundreds of instructions can be in-flight at any given moment, each of them operates on physical registers which are assigned just-in-time via a mechanism that maps the registers in the ISA to spare physical slots in very large banks. Each instruction is associated with a set of data that is entirely speculative for a very long time, including the instruction itself. 5/31

Instruction addresses, operands, dependencies are tracked as the CPU fetches and executes a stream of instructions which might or might not have to be executed, depending on branch prediction. In the case of a misprediction the state of the "wrong" instructions needs to be discarded. Similarly instruction timing is not predictable anymore. Memory accesses can take anything from a few cycles to hundreds, and fetch their data through different structures both inside and outside of the core. 6/31

Instruction faults cannot be predicted either: if a memory access fails because it access a protected memory area the flow of instructions must be stopped, undoing everything that should have happened before the faulting instruction and steering the core towards executing code provided by the operating system for such case, giving the impression that execution stopped right there in a perfectly sequential way. 7/31

And that's without mentioning the large amount of hidden state carried by a CPU purely for performance reasons: physical-to-memory address translations are done via tables stored in memory, but this data needs to be cached in a translation lookaside buffer inside the core. Cache lines can be shared by different cores and must track their state, are they owned by a core? Shared? Is the data dirty locally and needs to be fetched remotely? 8/31

The execution of every single instruction can alter a significant chunk of this enormous amount of state and must do so reliably. But as I mentioned before it's impossible to test all possible combinations and some sequences might lead to inconsistent or corrupted state, which in turn will manifest itself as a software bug. 9/31

Here's a few examples I've encountered: the instruction pointer is fetched from the stack while returning from a function call but it appears wrong, possibly because the wrong instruction pointer was sent to the instruction fetch pipeline: https://bugzilla.mozilla.org/show_bug.cgi?id=1746270 10/31

The code expected a piece of data to be loaded from memory, but the load/store unit returned stale data from a previous fetch: https://bugzilla.mozilla.org/show_bug.cgi?id=1687914 11/31

The instruction pointer associated with an instruction is corrupted, and what appears to be the currently executing instruction is not. You can tell because a load causes a store exception, or a jump causes an access exception: https://bugzilla.mozilla.org/show_bug.cgi?id=1820832 12/31

Other bugs could have even worse effects, such as AMD's infamous Barcelona TLB bug which would put the core from which recovery wasn't possible, effectively halting execution: https://arstechnica.com/gadgets/2007/12/linux-patch-sheds-light-on-amds-tlb-errata/ 13/31

In all these cases the likely culprit is a bug in the machinery that tracks the internal CPU state when an unlikely sequence of events happens: a rapid series of interrupt or context switches, execution timing of certain instructions while the processor leaves or enters a particular mode of execution. These are not unlikely software bugs where you missed checking a particular condition at a specific time, and most of the time it doesn't matter except for that one time when it does. 14/31

Reading the errata of any relatively recent CPU you will find the same wording applied to every known issue: "Under complex microarchitectural conditions...". That's hardwarese for "a state we had not anticipated we could end up in". Try looking it up yourself on an errata document such as this one: https://edc.intel.com/content/www/us/en/secure/design/confidential/products-and-solutions/processors-and-chipsets/tiger-lake/11th-generation-intel-core-processor-family-specification-update/errata-details/ 15/31

Now you might wonder if these kinds of bugs can be fixed after the fact. Well, sometimes they can, sometimes they can't. CPUs are not purely hard-coded beasts, they rely on microcode for part of their operation. Traditionally microcode is a set of internal instructions that the CPU ran to execute external instructions. That's mostly not the case anymore, and modern microcode ships not only with implementations of complex instructions but also a significant amount of configuration. 16/31

As an example microcode can be used to disable certain circuits. Imagine something like a loop buffer, a structure that captures decoded instructions and re-executes them in a loop bypassing instruction fetches. If it turns out to be buggy a microcode update might disable it entirely, effectively sacrificing an optimization for stability. 17/31

When implementing a new core it is commonplace to implement new structures, and especially more aggressive performance features, in a way that makes it possible to disable them via microcode. This gives the design team the flexibility to ship a feature only if it's been proven to be reliable, or delay it for the next iteration. 18/31

Microcode can also be used to work around conditions caused by data races, by injecting bubbles in the pipeline under certain conditions. If the execution of two back-to-back operations is known to cause a problem it might be possible to avoid it by delaying the execution of the second operation by one cycle, again trading performance for stability. 19/31

However not all bugs can be fixed this way. Bugs within logic that sits on a critical path can rarely be fixed. Additionally some microcode fixes can only be made to work if the microcode is loaded at boot time, right when the CPU is initialized. If the updated microcode is loaded by the operating system it might be too late to reconfigure the core's operation, you'll need an updated UEFI firmware for some fix to work. 20/31

But this is just logic bugs and unfortunately there's a lot more than that nowadays. If you've followed the controversy around Intel's first-generation Raptor Lake CPUs you'll know that they had issues that would cause seemingly random failures to happen. These bugs were caused by too little voltage being provided to the core under certain conditions which in turn would often cause a race condition within certain circuits leading to the wrong results being delivered. 21/31

To understand how this works keep this in mind: the maximum frequency at which a CPU can operate is dictated by the longest path through the circuits that make up a pipeline stage. Signals propagating via wires and turning transistors on and off take time, and because modern circuit design is strictly synchronous, all the signals must reach the end of the stage before the end of a clock cycle. 22/31

When a clock cycle ends, all the signals resulting from a pipeline stage are stored in a pipeline register. A storage element - invisible to the user - that separates pipeline stages. So if a stage adds two numbers for example, the pipeline register will hold the result of this addition. The next cycle this result will be fed to the circuits that make up the next pipeline stage. If the result of the addition I mentioned is an address for example, then it might be used to access the cache. 23/31

The speed at which signals propagate in circuits is proportional to how much voltage is being applied. In older CPUs this voltage was fixed, but in modern ones it changes thousands of times per second to save power. Providing just as little voltage needed for a certain clock frequency can dramatically reduce power consumption, but providing too little voltage may cause a signal to arrive late, or the wrong signal to reach the pipeline register, causing in turn a cascade of failures. 24/31