3 ms·

> But some applications are happy to trade absolute gflops for low latency and pay Intel out the nose for it. Can you provide one example of where Intel produc

> But some applications are happy to trade absolute gflops for low latency and pay Intel out the nose for it. Can you provide one example of where Intel products beat AMD's in latency in a way that prospective customers are willing to pay a premium for that performance delta?

loeg 8y ago
My employer, for example, probably can't live with the single thread perf regression that the combination of reduced IPC and steeper remote memory latency causes. We would love to pay less for CPUs, but they aren't the dominant cost in our platform (TBs of flash costs $$$$$) so the savings is marginal. (I don't really want to name my employer here, though.)
- rbanffy 8y ago
  Another case is memory. If your working dataset is, say, 24 TB, you want a machine that can have that much memory and you most likely have bigger problems than how much each of your 8 CPUs costs. With that kind of problem, you may want to abandon x86 altogether.
  - mmt 8y ago
    > With that kind of problem, you may want to abandon x86 altogether At this end of the spectrum, that means abandoning commodity hardware, with everything it implies, not just a change in binary architecture, as it might for ARM vs. x86 in end user devices. Recently, I looked up the max memory on IBM's system z, and it was only 32TB, but it looks like their Power Systems E980 can take 64TB. It appears that, like with their mainframe line, going that route means significant vendor lock-in, including proprietary memory modules. It's certainly interesting that the current generation of high-end x86 CPUs took a step backwards from the last generation, in that that their max is 12TB and just catching back up to 24TB isn't expected to happen until next year.
  - loeg 8y ago
    Fortunately, we do not have that problem :-).
    - rbanffy 8y ago
      I'd be extremely happy to have that kind of problem. For some reason, while I was in college, I got a complimentary subscription to a magazine called Supercomputing Review. It was one of the most interesting reads of that time.

loeg 8y ago

My employer, for example, probably can't live with the single thread perf regression that the combination of reduced IPC and steeper remote memory latency causes. We would love to pay less for CPUs, but they aren't the dominant cost in our platform (TBs of flash costs $$$$$) so the savings is marginal. (I don't really want to name my employer here, though.)

rbanffy 8y ago
Another case is memory. If your working dataset is, say, 24 TB, you want a machine that can have that much memory and you most likely have bigger problems than how much each of your 8 CPUs costs. With that kind of problem, you may want to abandon x86 altogether.
- mmt 8y ago
  > With that kind of problem, you may want to abandon x86 altogether At this end of the spectrum, that means abandoning commodity hardware, with everything it implies, not just a change in binary architecture, as it might for ARM vs. x86 in end user devices. Recently, I looked up the max memory on IBM's system z, and it was only 32TB, but it looks like their Power Systems E980 can take 64TB. It appears that, like with their mainframe line, going that route means significant vendor lock-in, including proprietary memory modules. It's certainly interesting that the current generation of high-end x86 CPUs took a step backwards from the last generation, in that that their max is 12TB and just catching back up to 24TB isn't expected to happen until next year.
- loeg 8y ago
  Fortunately, we do not have that problem :-).
  - rbanffy 8y ago
    I'd be extremely happy to have that kind of problem. For some reason, while I was in college, I got a complimentary subscription to a magazine called Supercomputing Review. It was one of the most interesting reads of that time.