4 ms·

Do they have meltdown and/or spectre bug?

deepnotderp 9y ago
They have spectre, any CPU with speculative execution and caches of any sort (excluding some capability machines, no ROP exploit machines and PLB enforced SASOS machines), including branch prediction (!) is vulnerable to spectre. (For those of you not in comp arch, that roughly translates to "literally every single non embedded CPU, and then some") That's what makes the flaw so deep and troublesome.
- skykooler 9y ago
  What about Atom? As I understood it Atom processors did not use speculative execution.
  - itsameta4 9y ago
    That was only the case until 2013 or so. Modern Atoms are vulnerable.
    - jandrese 9y ago
      And the old Atoms were so slow that few people bothered with them.
      - rasz 9y ago
        There were tens of millions of EeePCs sold. At one point Asus EeePCs were outselling all Macbooks.
  - rbanffy 9y ago
    IIRC, the first generation Xeon Phi had cores very similar to Atoms too.
    - craftyguy 9y ago
      Those cores were a bastard child of P54C[0] cores with the beginnings of (though not instruction compatible with) AVX512 + 64bit support (original Pentium had neither of these). They were not like Atom cores at all. The second generation Xeon Phi cores are pretty much based on beefed-up Airmont Atom cores. 0. https://en.wikipedia.org/wiki/P5_%28microarchitecture%29#P54C https://en.wikipedia.org/wiki/P5_%28microarchitecture%29#P54...
- iso-8859-1 9y ago
  Is the Mill architecture solving problems orthogonal to this?
  - thristian 9y ago
    The Mill doesn't do speculative execution, so I guess it should be safe from Spectre? It also doesn't have a normal TLB and a very different idea of memory protection, so perhaps it will be safe from Meltdown too. At the end of the day, though, Meltdown and Spectre are both consequences of all the caching between a modern CPU and main memory, so the Mill will likely need to at least be careful.
    - deepnotderp 9y ago
      Nah, the Mill does branch prediction, so it will still be in trouble. The Mill also does speculative execution, not OoOE but still speculative via predication.
- qwerty456127 9y ago
  Is speculative execution actually worth it? Can't compilers do the job and optimize execution sufficiently so we won't need speculative execution? Anyway, "caches of any sort" sounds frightening: I can remember disabling all the caches on Pentium-3 to make it about as slow as 80386 so some ancient games that can't run on fast CPUs would run ok and be playable.
  - thristian 9y ago
    Speculative execution is a workaround for the fact that CPUs read and process data much more quickly than main memory can provide it. The CPU tries to trace as many code-paths as possible so it can discover which memory addresses are likely to be used in the near future and queue them up so they'll be ready in the local cache when they're needed. Smarter compilers wouldn't really help, this is all about how the code is running for the next millisecond on this specific workload on this specific CPU, rather than how it's running today or on this cluster or in general.
  - AlphaSite 9y ago
    IA64 would like to reaffirm that this indeed worthwhile.
- ychen306 9y ago
  I thought the issue is that the fallback doesn't clear the cache affected by the speculation but not speculation per se?

They have spectre, any CPU with speculative execution and caches of any sort (excluding some capability machines, no ROP exploit machines and PLB enforced SASOS machines), including branch prediction (!) is vulnerable to spectre. (For those of you not in comp arch, that roughly translates to "literally every single non embedded CPU, and then some") That's what makes the flaw so deep and troublesome.

skykooler 9y ago
What about Atom? As I understood it Atom processors did not use speculative execution.
- itsameta4 9y ago
  That was only the case until 2013 or so. Modern Atoms are vulnerable.
  - jandrese 9y ago
    And the old Atoms were so slow that few people bothered with them.
    - rasz 9y ago
      There were tens of millions of EeePCs sold. At one point Asus EeePCs were outselling all Macbooks.
- rbanffy 9y ago
  IIRC, the first generation Xeon Phi had cores very similar to Atoms too.
  - craftyguy 9y ago
    Those cores were a bastard child of P54C[0] cores with the beginnings of (though not instruction compatible with) AVX512 + 64bit support (original Pentium had neither of these). They were not like Atom cores at all. The second generation Xeon Phi cores are pretty much based on beefed-up Airmont Atom cores. 0. https://en.wikipedia.org/wiki/P5_%28microarchitecture%29#P54C https://en.wikipedia.org/wiki/P5_%28microarchitecture%29#P54...
iso-8859-1 9y ago
Is the Mill architecture solving problems orthogonal to this?
- thristian 9y ago
  The Mill doesn't do speculative execution, so I guess it should be safe from Spectre? It also doesn't have a normal TLB and a very different idea of memory protection, so perhaps it will be safe from Meltdown too. At the end of the day, though, Meltdown and Spectre are both consequences of all the caching between a modern CPU and main memory, so the Mill will likely need to at least be careful.
  - deepnotderp 9y ago
    Nah, the Mill does branch prediction, so it will still be in trouble. The Mill also does speculative execution, not OoOE but still speculative via predication.
qwerty456127 9y ago
Is speculative execution actually worth it? Can't compilers do the job and optimize execution sufficiently so we won't need speculative execution? Anyway, "caches of any sort" sounds frightening: I can remember disabling all the caches on Pentium-3 to make it about as slow as 80386 so some ancient games that can't run on fast CPUs would run ok and be playable.
- thristian 9y ago
  Speculative execution is a workaround for the fact that CPUs read and process data much more quickly than main memory can provide it. The CPU tries to trace as many code-paths as possible so it can discover which memory addresses are likely to be used in the near future and queue them up so they'll be ready in the local cache when they're needed. Smarter compilers wouldn't really help, this is all about how the code is running for the next millisecond on this specific workload on this specific CPU, rather than how it's running today or on this cluster or in general.
- AlphaSite 9y ago
  IA64 would like to reaffirm that this indeed worthwhile.
ychen306 9y ago
I thought the issue is that the fallback doesn't clear the cache affected by the speculation but not speculation per se?