Nobody wnats low performance server processors?

By: David Kanter (dkanter.delete@this.realworldtech.com), March 11, 2016 10:38 am
Room: Moderated Discussions
Paul A. Clayton (paaronclayton.delete@this.gmail.com) on March 9, 2016 2:22 pm wrote:
> David Kanter (dkanter.delete@this.realworldtech.com) on March 8, 2016 8:50 pm wrote:
> [snip]
> > As it turns out, nobody really wants low performance server processors. Who
> > could have imagined? :)
>
> Even if you mean "processor" in the sense of socket/chip, I think "nobody" might be an exaggeration. The volume
> of potential sales might well be insufficient to support the development and manufacture of such systems, but
> I suspect there are use cases where the physical traits and incremental costs would be a good match.

Nobody is a bit hyperbolic, but if you cannot get enough customers to buy a lot you are dead. Remember that developing a processor probably runs around $100M. So you need enough volume to make everything work.

With 20M server processors, and 200M client processors, Intel gets the CPU core for free and only needs to spend money developing the cache, fabric, memory controller, PCIe, power management, etc.

> If you mean "low performance core", I am even more skeptical that "nobody" wants such. A core with
> 90% the performance of current high-end cores would be much smaller and measurably more energy efficient,
> especially if there is significant specialization for "server" workloads. While going below half performance
> may not be useful (given the area and power overheads of the memory system),

I agree that would be compelling. But so far, nobody (save IBM or Sun) has come up with a core that delivers much more than half the performance of an Intel server core. Take Cavium, they appear to have comparable performance to a very mid-range Xeon - with 2-3X the core count and higher power consumption than a 2650L.

The highest IPC core is Apple's, but it's limited to fairly low frequency and might not do particularly well at server workloads (since its not designed for those). Moreover, even having a core isn't enough. Just look at AppliedMicro. Their first design may have had a great core, but it didn't matter because the memory controller was unimpressive.

> It also depends on what performance one is measuring. A wimpy core that can only achieve 25% (or less) of the
> DGEMM performance of a high performance core may achieve 90% of the performance on a low ILP workload.

That's true, and if you are just running one workload that is really about PCIe, then it's quite possible there are better solutions.

The strength of Intel's cores is that they tend to run everything well. And in many cases, they are simply the best.

THe other thing to remember is that performance matters, but so does power. People buy servers typically based around TCO and operational cost really matters.

I've seen calculations indicating that AMD would have to give people money to use their server chips given the performance/power profile.

David
< Previous Post in ThreadNext Post in Thread >
TopicPosted ByDate
X-Gene 3 power enelopeRonald Maas2016/03/07 10:23 PM
  X-Gene 3 power enelopeDavid Kanter2016/03/08 09:50 PM
    X-Gene 3 power enelopecanon2016/03/09 09:32 AM
      X-Gene 3 power enelopeAlberto2016/03/09 03:38 PM
    Nobody wnats low performance server processors?Paul A. Clayton2016/03/09 03:22 PM
      Nobody wnats low performance server processors?Ricardo B2016/03/09 05:05 PM
        Nobody wants low performance server processors?Doug S2016/03/10 09:03 AM
          Nobody wants low performance server processors?Ricardo B2016/03/10 10:35 AM
            bad size comparison between Apple A9 and skylake coresHeikki Kultala2016/03/12 05:37 AM
              bad size comparison between Apple A9 and skylake coresAlberto2016/03/12 02:04 PM
                bad size comparison between Apple A9 and skylake coresMaynard Handley2016/03/13 02:39 PM
                  bad size comparison between Apple A9 and skylake coresDoug S2016/03/13 06:10 PM
                  Comparing A9 and SkylakeDavid Kanter2016/03/13 06:15 PM
                    Core sizes & performanceHans de Vries2016/03/14 05:36 AM
                      Core sizes & performanceDavid Kanter2016/03/14 09:16 AM
                        Core sizes & performanceVertexMaster2016/03/14 12:37 PM
                          Core sizes & performancejuanrga2016/03/15 05:40 AM
                          Core sizes & performanceDavid Kanter2016/03/15 06:59 PM
                            Core sizes & performanceyellowhighlighter2016/03/15 09:30 PM
                      Core sizes & performance: Zen & K12juanrga2016/03/15 05:54 AM
        Common wisdom??Paul A. Clayton2016/03/10 09:33 AM
          Common wisdom??Ricardo B2016/03/10 10:49 AM
            Common wisdom??wumpus2016/03/10 01:32 PM
        Nobody wnats low performance server processors?juanrga2016/03/12 12:43 PM
      Nobody wnats low performance server processors?Linus Torvalds2016/03/10 11:52 AM
      Nobody wnats low performance server processors?David Kanter2016/03/11 10:38 AM
        Nobody wnats low performance server processors?MrJava2016/03/11 04:52 PM
        Nobody wnats low performance server processors?wumpus2016/03/12 10:15 AM
        Nobody wnats low performance server processors?nobody in particular2016/03/12 11:27 AM
          Nobody wnats low performance server processors?David Kanter2016/03/13 01:04 AM
  X-Gene 3 power enelopejuanrga2016/03/12 12:33 PM
    X-Gene 3 power enelopeRonald Maas2016/03/14 07:23 PM
      X-Gene 3 power enelopeGabriele Svelto2016/03/15 02:37 AM
        X-Gene 3 power enelopecanon2016/03/15 10:30 AM
        X-Gene 3 power enelopeMichael S2016/03/15 10:47 AM
          X-Gene 3 power enelopecanon2016/03/15 12:03 PM
        X-Gene 3 power enelopeRonald Maas2016/03/15 07:39 PM
      X-Gene 3 power enelopedmcq2016/03/16 07:28 AM
        X-Gene 3 power enelopeanon2016/03/17 07:44 PM
          X-Gene 3 power enelopeRonald Maas2016/03/22 07:49 AM
Reply to this Topic
Name:
Email:
Topic:
Body: No Text
How do you spell purple?