Cloudflare CEO: Why we’re switching to ARM-based servers
twitter.com
twitter.com
Most powerful Xeon is the 28 core Platinum 2180 at $10k RSP and >200W TDP. Due to the nature of Intel turbo boost it almost always operates at TDP under load. Our ARM is the Centriq 2452 at less than $1400 RSP, 46 cores and 120W TDP, that it never hits. Beats Xeon 9/10 workloads.
I just hope that thanks to large companies like CloudFlare buying large numbers of ARM servers their prices finally start going down to a reasonable level.
There is a big chance, that whatever cloudflare is doing on that machine, is massive parallel and simple.
It is taking decades at this rate. Or at least a decade.
Not to long ago when Intel was yanking the rug out from under the SPARC and POWER and MIPS and $RANDOM_PROPRIETARY_MAINFRAME_CISC datacenter marketshares.
Well it took decades...
To be honest I am very much surprised at how cheap Qualcomm are selling their chip.
This is going to be the start of avalanche. Once the software improves and make a self sustainable ecosystem, everything else will come. I still dont believe Qualcomm could achieve their 5% Server target by 2020 goal. But it is still a good start in longer terms success.
What is more worrying is Intel's growth and Cash Cow is now at risk in the mid term. It will be interesting to see how Intel react to this. There is serious price pressure from AMD EPYC and Qualcomm Centriq. Intel DC 2018 may ship lower unit YoY and lower revenue / unit YoY. PC market is being attacked by AMD, while not at risk, Intel is definitely earning less as they increase core count.
Personally I would have liked to see POWER9 or RISC-V in server space, but both still missing in action.
AMD: I need some threadripper or server Zen motherboards, again there's 4+ OEMs with many designs.
Both amd and Intel: all sorts of stuff available that meets open compute platform dimensions for motherboards. Or for ordinary 1U servers, or atx size boards that can be mounted in DIY blade enclosures.
ARM: where are the motherboards I can buy right now in small to low quantities, and what are their specs/price? Who makes them? It I want to order 40 or 400 tomorrow... where do I go?
Edit: I just spent ten minutes googling for arm centriq platform motherboards and cannot find a single product I can actually buy. As compared to how easily I could start the process of ordering a hundred 1U xeon based servers in the next half hour.
At a massive scale such as cloudflare, sure you can afford to custom order thousands of motherboards. But for a lot of operators who are not nearly as big as them, the existing economies of scale in the x86-64 platform are a greater advantage. Right now I'd be willing to bet that dual socket current generation xeon motherboards are manufactured in much greater numbers, as much as a ratio of 500:1, compared to ARM server motherboards. Being able to mix and match whitebox parts from Taiwan and China sources to DIY bare metal blade hypervisors is a huge advantage.
I'm constantly amazed by people wanting to see Intel die for a minor price decrease. The open source drivers for intel gpus might have not been provided by intel but at the very least they exist. The same cannot be said by ARM. When it comes to proprietary software ARM is at least as evil as intel. By replacing intel with ARM you're going from bad to worse.
Does anyone know if you can get hold of a Centriq cpu & compatiple motherboard to build a local server or two? I fancy one in our lab rack to play around with.
I'm assuming that this is not possible and that they are currently just for datacentre partners who make a large order. But would be great to hear otherwise.
One of the workloads that Cloudflare does is encryption, and RISC-V has been planning a vector extension that can be reconfigured, so that the jump from 128-bit to 256-bit native execution is not an ISA change and doesn't require recompiling.
AMD offers 256 bit vector instructions but they take twice as long because the Zen cores can only process 128 bit at a time.
I'm really curious at what the blog post will reveal on this.
1. A lot of cloud services are abstracting CPU architecture. E.g. you don't care on which CPU AWS S3 runs.
2. ARM is using less power than x86.
3. ARM license fees are so low comparing to Intel high gross margins. Whole mobile world runs on ARMs and little profit is made on CPUs there comparing to Intel cash cow server chips.
If you use a storage apppiance - be it a local NAS or S3 or B2 - you likely aren’t interested in the cpu running it.
A better comparison would be if you didn’t care what cpu Amazon Lambda ran on. Which, maybe you don’t, if you use it with an interpreted language.