The patch for meltdown results in a performance hit of between 2% and 20%. It’s hard to pin down an exact number because it varies both by CPU and use case.
It basically varies from chip to chip, and program to program.
Speculative execution is when a program hits some kind of branch (like an if-then statement) and the CPU just goes ahead and calculates as if it’s true, and progresses down that line until it learns “oh wait it was false, just scrub all that work I did so far down this branch.” So it really depends on what that specific chip was doing in that moment, for that specific program.
It’s a very real performance boost for normal operations, but for cryptographic operations you want every function to perform in exactly the same amount of time, so that something outside that program can’t see how long it took and infer secret information.
These timing/side channel attacks generally work like this: imagine you have a program that tests if variable X is a prime number, by testing if every number smaller than X can divide evenly, from 2 on to X. Well, the bigger X is, the longer that particular function will take. So if the function takes a really long time, you’ve got a pretty good idea of what X is. So if you have a separate program that isn’t allowed to read the value of X, but can watch another program operate on X, you might be able to learn bits of information about X.
Patches for these vulnerabilities changes the software to make those programs/function in fixed time, but then you lose all the efficiency gains of being able to finish faster, when you slow the program down to the weakest link, so to speak.
How much slower is s CPU without this functionality built in?
The patch for meltdown results in a performance hit of between 2% and 20%. It’s hard to pin down an exact number because it varies both by CPU and use case.
It basically varies from chip to chip, and program to program.
Speculative execution is when a program hits some kind of branch (like an if-then statement) and the CPU just goes ahead and calculates as if it’s true, and progresses down that line until it learns “oh wait it was false, just scrub all that work I did so far down this branch.” So it really depends on what that specific chip was doing in that moment, for that specific program.
It’s a very real performance boost for normal operations, but for cryptographic operations you want every function to perform in exactly the same amount of time, so that something outside that program can’t see how long it took and infer secret information.
These timing/side channel attacks generally work like this: imagine you have a program that tests if variable X is a prime number, by testing if every number smaller than X can divide evenly, from 2 on to X. Well, the bigger X is, the longer that particular function will take. So if the function takes a really long time, you’ve got a pretty good idea of what X is. So if you have a separate program that isn’t allowed to read the value of X, but can watch another program operate on X, you might be able to learn bits of information about X.
Patches for these vulnerabilities changes the software to make those programs/function in fixed time, but then you lose all the efficiency gains of being able to finish faster, when you slow the program down to the weakest link, so to speak.