juanrga (noemail.delete@this.juanrga.com) on October 11, 2018 11:34 am wrote:
> anon (spam.delete.delete@this.this.spam.com) on October 11, 2018 7:08 am wrote:
> > juanrga (noemail.delete@this.juanrga.com) on October 11, 2018 3:11 am wrote:
> > > nobody in particular (nobody.delete@this.nowhe.re) on October 10, 2018 8:55 am wrote:
> > > > More info here
> > > >
> > > > "In essence, requests per second (RPS) per Watt is a critical metric that Qualcomm’s
> > > > ARM64 46 core Falkor chip had a big advantage over Intel’s Skylake 4116. Embracing
> > > > the value of optionality and market competition, we made some noise.
> > > >
> > > > Intel proposed to co-innovate with us an off-roadmap 24-core
> > > > Xeon Gold CPU specifically made for our workload
> > > > offering considerable value in Performance per Watt. For this generation, we continue using Intel as system
> > > > solutions are widely available while we’re working on realizing ARM64’s benefits to production."
> > > >
> > > > Looks like Centriq was far less of a done deal for Cloudflare than many assumed.
> > >
> > > The blog claims custom Xeon "offering considerable value in Performance per Watt".
> > >
> > > Then adds that the custom Xeon increases "RPS by 200% from doubling the amount of cores" and increases
> > > "the power consumption by 174% from increasing the CPUs TDP from 85W to 150W each".
> > > Yes the number of cores is doubled, but the frequency drops
> > > from 2.1GHz to 1.9GHz. So the RPS is only 181%, not "200%".
> > >
> > > So performance per watt is only 4% better, and far from the 2x gap offered by Falkor.
> >
> > If it's better at all.
> >
> > Cloudflare doesn't care about efficiency.
>
> Ok, they can just claim that they care about density or that they got a rebate couldn't
> resist or whatever it is the reason, but then they could stop claiming that the custom Xeon
> is "offering considerable value in Performance per Watt" compared to the old Xeon.
>

Yeah, but that doesn't sound great.
The honest conclusion of the Centriq evaluation would've been "46 threads are about as fast as 48 threads. 1400$ is less than 2x1000$. This appears to be cheaper".
And now the conclusion would be "96 threads are faster than 46 threads, even at lower clockrate. One server costs less than two, even if the CPUs are less efficient. Buying half as many servers appears to be cheaper."

If you buy something because it's more efficient you're doing it for the environment, for everyone, and should be commended. If you buy something because it's cheaper you're just doing what any profit-oriented company should do. Being honest and admitting that it's less efficient but you don't really care doesn't get you a whole lot of goodwill.

The whole post can summed up with "Intel offered us downclocked, cut down to Xeon Gold features (or less) 24 core SKUs for 2xxx$. Qualcomm doesn't offer dual socket so they can't compete with that. We'll just keep building the exact same servers, but with twice the cores now. Try again next year Qualcomm".
But that's not a good blog post. Doing the same that you were already doing but now with more cores? That's what everybody has been doing for years. That's not revolutionary. You can't brag about that.

> > There is a reason why they
> > had only one power consumption/efficiency test in their evaluation.
> > They care about cost and density and they only care about density because it affects cost.
> > Single socket 46 thread Falkor vs dual socket 48 thread Skylake is a win for Falkor.
> > Similar performance, lower power consumption, lower cost per performance.
> > Against dual socket 96 thread Skylake (at Xeon Gold prices) the density
> > and system cost advantage is so large that efficiency becomes irrelevant.
>
>