3 ms·

I do prefer Intel's solution of slowing down over AMD's solution of crashing. I guess the good news is if this does turn out to be a power delivery related bug,

• bdd 7y ago
  Maybe it wasn't very clear in my comment but "correctness issues" more-or-less literally mean the instruction's result is incorrect. Like you ask 2*2 and it gives you 0xfff1309f. So depending on how you rely on that result, you may crash. Extending on my examples, bad RSA multiplication means failed authentication. Bad AES construction means encrypting things to garbage, impossible to decrypted by the recipient. So, let's say if you rely on a result from such instruction to do memory references, you're definitely going to crash.
• olliej 7y ago
  Slowing down and producing the correct answers is clearly superior to crashing and/or incorrect answers. I still don’t understand what it is about avx2 that results in these kinds of issues - is it really just a matter of increased number of execution units running at once causing weird power and heat issues?
  • wtallis 7y ago
    > is it really just a matter of increased number of execution units running at once causing weird power and heat issues? That's been the root cause of all the previous AVX weirdness I'm aware of. It's probably the case for these Threadrippers, too. Modern CPUs have very high power density and run at pretty low voltage, which results in insane current delivery requirements.
  • paulmd 7y ago
    Not increased number of execution units, but increased number of transistors, yeah. A normal multiply does 1 number at a time (32 bit for simplicity). AVX2 can do up to 8 multiplications at a time. That's a huge amount more transistors firing all at once and that causes the voltage to start to droop. AVX-512 takes that even further and now it's 16 multiplications per unit, oh and Intel moved from 1x256-bit unit per core on Haswell/Broadwell to 2x512-bit units on Skylake-X, so it's potentially 32 multiplications at a time - 4x as much as AVX2. Basically to prep for that much power being drawn all at once, the chip has to switch to a higher-voltage mode to account for the voltage droop caused by all those transistors switching at once in one place. It takes time for the regulator (on-chip Fully Integrated Voltage Regulator or motherboard VRM) to wind up the voltage far enough. https://travisdowns.github.io/blog/2020/01/17/avxfreq1.html https://travisdowns.github.io/blog/2020/01/17/avxfreq1.html https://travisdowns.github.io/assets/avxfreq1/fig-volts256-1.svg https://travisdowns.github.io/assets/avxfreq1/fig-volts256-1... At this level behavior is intensively analog and thermals/voltage both significantly affect transistor current draw and switching time, which feeds back into thermals and power consumption/voltage droop. This gets even more problematic on 7nm/10nm type nodes and especially in GPUs where you are doing a huge amount of vector arithmetic all the time. Essentially it is no longer possible to design processors that are 100% stable under all potential execution conditions, or even under normal operating conditions, so you have to have power watchdog circuitry that realizes when it's getting close to brownout/missing its timing conditions and slows itself down to stay stable. That's why AMD introduced clock stretching in a big way with Zen2 (despite the fact that it's nominally been around since Steamroller). NVIDIA piloted this with Pascal, AMD piloted it with Vega and brought it to CPU with Zen2. You simply cannot design the processor to be 100% stable at competitive clocks anymore. You have to have power management that's smart enough to withstand small transient power conditioning faults. https://semiengineering.com/managing-voltage-drop-at-107nm/ https://semiengineering.com/managing-voltage-drop-at-107nm/ https://semiengineering.com/power-delivery-affecting-performance-at-7nm/ https://semiengineering.com/power-delivery-affecting-perform... https://www.realworldtech.com/steamroller-clocking/ https://www.realworldtech.com/steamroller-clocking/