By: anon (anon.delete@this.anon.com), October 3, 2012 6:50 pm
Room: Moderated Discussions
Robert Myers (rbmyersusa.delete@this.gmail.com) on October 3, 2012 10:52 am wrote:
> David Kanter (dkanter.delete@this.realworldtech.com) on September 18, 2012 12:26
> pm wrote:
> > Our latest article has just gone online:
> >
> >
> "Near-threshold voltage computing
> > extends the voltage scaling associated
> with Moore's Law and dramatically
> > improves power and energy efficiency.
> The technology is superb for throughput,
> > at the cost of latency, and best
> suited to Intel's products for HPC and mobile
> > graphics."
> >
>
> I've
> sort of dropped this subject because I am busy with other things, but HPC
> workloads are not a uniform universe. As I have been repeatedly informed,
> latency is still a driving issue for many HPC workloads. I'd make the trade of
> latency vs. power, computational density, and/or bandwidth any day, but I don't
> know how many HPC users would agree.
What do you mean by latency?
There is the old problem with how many physical simulations are scaled to many cores, which is that the speed of the result is determined by the speed of the slowest job. So in a 1000 core system, a core which is taking a periodic interrupt or administrative job that causes its job to run 1% slower, is better off not being used at all, because it effectively costs 10 cores of throughput.
This is the latency issue that I know HPC cares about. But if they could choose to have a CPU which is 200% the throughput but with a 10% variation between CPUs, surely there is no downside to that (provided that perf/watt stays the same and overall power draw scales with the idle time introduced).
> David Kanter (dkanter.delete@this.realworldtech.com) on September 18, 2012 12:26
> pm wrote:
> > Our latest article has just gone online:
> >
> >
> "Near-threshold voltage computing
> > extends the voltage scaling associated
> with Moore's Law and dramatically
> > improves power and energy efficiency.
> The technology is superb for throughput,
> > at the cost of latency, and best
> suited to Intel's products for HPC and mobile
> > graphics."
> >
>
> I've
> sort of dropped this subject because I am busy with other things, but HPC
> workloads are not a uniform universe. As I have been repeatedly informed,
> latency is still a driving issue for many HPC workloads. I'd make the trade of
> latency vs. power, computational density, and/or bandwidth any day, but I don't
> know how many HPC users would agree.
What do you mean by latency?
There is the old problem with how many physical simulations are scaled to many cores, which is that the speed of the result is determined by the speed of the slowest job. So in a 1000 core system, a core which is taking a periodic interrupt or administrative job that causes its job to run 1% slower, is better off not being used at all, because it effectively costs 10 cores of throughput.
This is the latency issue that I know HPC cares about. But if they could choose to have a CPU which is 200% the throughput but with a 10% variation between CPUs, surely there is no downside to that (provided that perf/watt stays the same and overall power draw scales with the idle time introduced).
