Cloudflare CPU has better perf/W

By: anon (spam.delete.delete@this.this.spam.com), October 15, 2018 11:52 am
Room: Moderated Discussions
David Kanter (dkanter.delete@this.realworldtech.com) on October 15, 2018 11:09 am wrote:
> anon (spam.delete.delete@this.this.spam.com) on October 15, 2018 9:10 am wrote:
> > David Kanter (dkanter.delete@this.realworldtech.com) on October 15, 2018 8:39 am wrote:
> > >
> > > > Well, exactly! "Co-innovating" a custom CPU "specifically made for our workload offering considerable value
> > > > in Performance per Watt" as their blog post says on one hand and a throwaway bin whose perf/W is worse than
> > > > everything Intel offers on the other, which is evident from figures, are clearly not the same thing.
> > >
> > > The custom SKU mentioned by cloudflare actually has better perf/w than their prior generation. You should
> > > redo your math, it's about 5% better and also boosts system perf/w with more cores per system.
> > >
> > > David
> >
> > Prior generation as in Broadwell or the Xeon Silver?
>
> Their comparison was of G8 vs. G9. And you have to consider the whole
> server, not just the CPU. Doubling the CPUs/server is quite valuable.
>

Well their comparision never showed actual benchmarks.
Either way the point was the the CPU isn't more efficient and indeed it is not.
Not just some, but all of the value comes from doubling the cores per server.

The whole article is a PR/marketing piece.

It starts with:
"We expect this CPU to perform with better RPS per Watt right off the bat; increasing the RPS by 200% from doubling the amount of cores, and increasing the power consumption by 174% from increasing the CPUs TDP from 85W to 150W each."

Then it shows a server with twice the cores at lower clockrate doing 2.3 times the requests at 90% instead of 100% load.
If they haven't compared it to the Broadwell 10 core servers then a significant part of this must be due to the better SSDs, faster networking and 33% more, higher clocked RAM with 25% less bandwidth, in which case we haven't really learned anything about the CPU or Skylake just magically becomes 25% faster by lowering the clockrate and adding more cores, in that case it's straight up magic.
In fact the new server is so good it only needs half the read IOPS for more than twice the requests.
So they add 33% more RAM, then double the requests and their RAM cache misses per second go down?
Quite the magic CPU.

Afterwards comes their conclusion:
"G9 is actually 1.5x more power efficient than G8. Temperature readings were checked as well. Inlet and outlet chassis temps, as well as CPU temps, are well within operating temperatures.

Now that’s muscle! In other words for every 3 G8 servers, just 2 of those G9's would take on the same workload."

So it does 2.3x the requests at lower CPU load and latency but nope, can't replace 2 old servers with 1 new server, they are only 50% more efficient, so we need 2 to replace 3. What? That's not how it works. This is all just marketing bla bla "see, efficiency is important". Even if they showed exactly the same efficiency they'd still buy them because it gets them more than twice the performance.

They could've bought similar CPUs any time. The whole server offers better performance per Watt, what the CPUs offer is more cores per server/dollar.


> > It's slightly better than the Xeon Silver but 2.5% doesn't look like "considerable value in
> > Performance per Watt" to me. At least it's nowhere near the ~100% Centriq would get them.
>
> > They claim RPS scale with the number of cores, completely
> > ignoring the clockrate and I have a hard time believing
> > that. They're not that memory bound that a 10% change in clockrate doesn't affect performance at all.
>
> They definitely aren't memory bound...they are only using 4 channels of memory.
>

Exactly.

> > And the math checks out, naively going by rated clockrates and
> > TDP that custom SKU is less efficient than any 24 core list SKU.
>
> There is only one 24C - the Xeon 8160. So yes, they are basically buying a slower/less
> efficient version of the 8160...but probably for a lot less money.
>
> David

Does the 8168 not count?
< Previous Post in ThreadNext Post in Thread >
TopicPosted ByDate
Cloudflare jumps to custom Intelnobody in particular2018/10/10 07:55 AM
  Cloudflare jumps to custom Intelnone2018/10/10 08:25 AM
  Cloudflare jumps to custom Intelanon2018/10/10 08:39 AM
  Cloudflare jumps to custom IntelMaynard Handley2018/10/10 09:26 AM
    Cloudflare jumps to custom Intelnobody in particular2018/10/10 12:03 PM
      Cloudflare jumps to custom IntelWilco2018/10/10 12:25 PM
        Cloudflare jumps to custom Intelanon2018/10/10 01:07 PM
          Cloudflare jumps to custom Intelhobold2018/10/10 01:14 PM
          Cloudflare jumps to custom IntelWilco2018/10/10 01:58 PM
            Cloudflare jumps to custom Intelanon2018/10/10 07:56 PM
              Cloudflare jumps to custom IntelWilco2018/10/11 01:18 AM
                Cloudflare jumps to custom IntelMichael S2018/10/11 02:43 AM
                Cloudflare jumps to custom Intelanon2018/10/11 02:30 PM
          Cloudflare jumps to custom IntelBeastian2018/10/10 04:33 PM
            Cloudflare jumps to custom IntelDoug S2018/10/11 01:38 AM
          Qualcomm restructured server chip unitjuanrga2018/10/11 02:04 AM
            Qualcomm restructured server chip unitnone2018/10/11 03:38 AM
              Qualcomm restructured server chip unitMichael S2018/10/11 06:48 AM
                Qualcomm restructured server chip unitnone2018/10/11 08:15 AM
      Cloudflare jumps to custom Intelanon2018/10/10 12:29 PM
  Cloudflare jumps to custom IntelDoug S2018/10/10 11:10 AM
    Cloudflare jumps to custom Inteldmcq2018/10/10 11:27 AM
    Cloudflare jumps to custom Intelanon2018/10/10 02:07 PM
      Cloudflare jumps to custom IntelMichael S2018/10/11 03:06 AM
        Cloudflare jumps to custom Intelanon2018/10/11 06:02 AM
          Why can't two "computers" be on one board?Paul A. Clayton2018/10/11 06:49 PM
            Why can't two "computers" be on one board?rwessel2018/10/12 12:01 AM
              TwinBladesPaul A. Clayton2018/10/12 04:52 AM
            Why can't two "computers" be on one board?Doug S2018/10/12 01:01 AM
              Why can't two "computers" be on one board?Etienne2018/10/12 02:04 AM
              [O.T.] bladesMichael S2018/10/12 02:17 AM
                [O.T.] bladesAdrian2018/10/12 04:17 AM
              Why can't two "computers" be on one board?anon2018/10/12 03:04 AM
                Why can't two "computers" be on one board?rwessel2018/10/12 07:29 AM
                  Why can't two "computers" be on one board?anon2018/10/12 07:55 AM
              1S vs 2SWes Felter2018/10/12 01:30 PM
                1S vs 2SDoug S2018/10/12 02:35 PM
            Why can't two "computers" be on one board?anon2018/10/12 03:10 AM
            Why can't two "computers" be on one board?hobel2018/10/12 06:53 AM
      Cloudflare jumps to custom Intelanon.12018/10/11 09:16 AM
  Weird numbersjuanrga2018/10/11 02:11 AM
    Weird numbersanon2018/10/11 06:08 AM
      Weird numbersjuanrga2018/10/11 10:34 AM
        Weird numbersanon2018/10/11 11:24 AM
  Look at the figures, folksAM2018/10/11 08:15 AM
    Look at the figures, folksFoo_2018/10/11 09:22 AM
      Look at the figures, folksAM2018/10/11 10:02 AM
        Look at the figures, folksanon2018/10/11 02:49 PM
          Look at the figures, folksSimon Farnsworth2018/10/12 12:19 AM
            Some power efficiency figuresAM2018/10/12 03:59 AM
              Some power efficiency figuresSimon Farnsworth2018/10/12 09:00 AM
                Some power efficiency figuresMaynard Handley2018/10/12 09:33 AM
                  Some power efficiency figuresSimon Farnsworth2018/10/12 10:51 AM
                    Some power efficiency figuresanon2018/10/12 12:55 PM
                    Some power efficiency figuresMaynard Handley2018/10/12 02:37 PM
                      Some power efficiency figuresanon2018/10/12 03:04 PM
                        Some power efficiency figuresMaynard Handley2018/10/12 03:16 PM
                          Some power efficiency figuresanon2018/10/13 01:13 AM
                          Some power efficiency figuresanon2018/10/13 04:06 AM
                      Some power efficiency figuresSimon Farnsworth2018/10/13 03:52 AM
                        Some power efficiency figuresanon2018/10/13 04:17 AM
                Reread what you saidAM2018/10/14 08:14 AM
          Important clue for several postersAM2018/10/12 03:48 AM
            Important clue for several postersjuanrga2018/10/12 09:07 AM
              Important clue for several postersWilco2018/10/12 12:50 PM
                Thanks (NT)juanrga2018/10/13 03:17 AM
            Important clue for several postersWes Felter2018/10/12 01:42 PM
              Important clue for several postersAM2018/10/14 08:00 AM
                Cloudflare CPU has better perf/WDavid Kanter2018/10/15 07:39 AM
                  Cloudflare CPU has better perf/Wanon2018/10/15 08:10 AM
                    Cloudflare CPU has better perf/WDavid Kanter2018/10/15 10:09 AM
                      Cloudflare CPU has better perf/Wanon2018/10/15 11:52 AM
            Important clue for several postersanon2018/10/13 01:20 AM
              Important clue for several postersAM2018/10/14 07:58 AM
                Important clue for several postersanon2018/10/14 02:29 PM
                  Important clue for several postersAM2018/10/16 09:18 AM
                    Important clue for several postersanon2018/10/16 10:35 AM
                    Important clue for several postersanon2018/10/17 03:01 PM
    Look at the figures, folksDavid Kanter2018/10/13 05:08 AM
      Look at the figures, folksAM2018/10/14 07:50 AM
        Look at the figures, folksDavid Kanter2018/10/14 08:25 PM
          Look at the figures, folksMichael S2018/10/15 01:16 AM
            What?!AM2018/10/16 09:12 AM
          ARM servers in AzureMichael S2018/10/15 01:19 AM
            ARM servers in Azurejuanrga2018/10/15 04:41 AM
              ARM servers in AzureMichael S2018/10/15 04:46 AM
              ARM servers in AzureAnon2018/10/15 04:54 AM
                ARM servers in Azureanon2018/10/15 06:40 AM
          Look at the figures, folksAM2018/10/16 09:07 AM
        Look at the figures, folksMichael S2018/10/15 01:25 AM
        What figures?someone2018/10/15 04:35 AM
          What figures?AM2018/10/16 09:03 AM
            What figures?someone2018/10/17 07:19 AM
              What figures?wumpus2018/10/18 07:46 AM
      Look at the figures, folksJames2018/10/15 04:16 AM
        Look at the figures, folksAnne O. Nymous2018/10/15 04:31 AM
          Look at the figures, folksanonymou52018/10/15 04:47 AM
            Look at the figures, folksAnne O. Nymous2018/10/15 04:55 AM
        Watch the financial reportsAnon2018/10/15 05:00 AM
Reply to this Topic
Name:
Email:
Topic:
Body: No Text
How do you spell green?