Is Facebook Planning a Move to ARM Based Servers? | Hacker News Reader

Is Facebook Planning a Move to ARM Based Servers?

josephscott.org

27 pointsjosephscott22 comments

So is Google:

http://www.bloomberg.com/news/2013-12-12/google-said-to-mull...

The proposed applications (network control plane) don't correspond to what you would write in PHP (the main Facebook app).

The arm64 stuff on the other hand is looming like being the first new RISC architecture since er alpha and there may well be a move towards using it. There was a story recently that Google has an ARM license too.

Why ARM? What are the benefits of ARM compared to other architectures? Are there drawbacks to the architecture?

For the same reason Apple is making its own custom ARM CPU right now - it can be highly integrated with whatever they want it to be, and it has exactly what they need, and none of the stuff they don't need. You can build solutions that are much more custom with ARM.

Isn't it cheaper just to use a general purpose CPU than design and fabricate a custom one?

At a given volume, a custom CPU is a commodity CPU.

What volume is that for the type of ARM chip Facebook might be using?

Google is reckoned to have about 2 million servers[1] and they're probably dual- or quad-processor. I guess FB have ambitions at that scale.

The expensive bit of running a datacentre is power, both for the machines themselves, and for cooling, and these are linked. For every Watt you put in in electricity, you have to take it out again as heat. It costs more than the real estate and more than payroll. Savings here can add up very quickly.

[1] https://plus.google.com/app/basic/stream/z13ttdlibtq5tjv5e04...

It doesn't take that many devices too offset design cost.

Well that's not a property of ARM the CPU but ARM the business model (IP licensing). And ARM aren't the only player in this space, there is also MIPS (and others).

ARM has the momentum, compared to MIPS.

Sure, but there was a time that MIPS was in some of the most powerful machines in the world, from workstations to supercomputers. Things can change pretty fast in this industry.

While ARM has a lot more momentum I wouldn't completely discount MIPS. They are still out there and churning away in places you wouldn't expect. If I remember correctly the PIC32 is nothing but a modified MIPS core.

In addition, China has staked its domestic chip design efforts on MIPS64:

http://en.wikipedia.org/wiki/Loongson

Incidentally, I think the following sentence from the Wikipedia introduction is misleading / untrue.

  China uses GNU/Linux with its Loongson processor family
  to achieve technology independence.

It reads to me like a statement that Linux is widely used throughout China, whereas my understanding is that use of non-Microsoft OSs is basically negligible in China[1]. Perhaps someone who knows more than me about this area would like to remove it.

[1] http://gs.statcounter.com/#os-CN-monthly-201211-201311

OpenSPARC is another one in this space. I was working on a wifi daughter board for a project a while ago that had a SPARC at its heart. It really made sense for them. They were already in the business of fabing their own complicated ICs, they didn't need a really power CPU and power wasn't a huge concern to them. So they got a pretty decent CPU with mature tooling without the licensing to ARM.

These are only my guesses at their logic though. I was just an outsider looking in at a closed black box.

And IBM with PPC.

I think if MIPS were going to get in this space (which really they should have been trying to do, instead of chasing the mobile tail) they would have succeeded already, since things like the Cavium chips seem closer to what a lot of these people really want than anything ARM had until recently.

The end of MIPS as an independent company has definitely created huge questions about the future viability of the architecture. If I understand the deal right Imagination (of PowerVR) got the day-to-day company, and a holding group, including ARM, got access to the design patents. That means future ARM chips are likely to incorporate more of whatever advantages the MIPS teams had been hoping to get.

I think the only certainty in all this is Intel are going to be upset one way or another. ARM can simply redefine what people mean by ARM architecture every couple of years, whereas x86 does not move so quickly. Intel will either end up fabbing ARM designs, which is a way to expose just how much they've relied on (great) fab tech to stay ahead, or they'll stick to the x86 guns and slowly but surely the combined mass of everyone else in the semiconductor business will chip them away.

As above; ARM grew from a very low-power beginning (some 250mW total for the original ARM-2 chipset of CPU, IOC, MEMC, and VIDC, running at 8MHz on 2 micron tech, in 1989), and has remained low-power since. Whilst the CPU’s obviously only a part of any server, I’d imagine that the savings in power usage across as many servers as companies like Google and Facebook have are the driving force behind these evaluations.

They've been hiring for this for a while now, https://www.facebook.com/careers/department?dept=engineering...

Didn't HP consider making blade like units with up to 20 servers per unit? I saw that they canceled that due to some deal with Intel.

Maybe for things like photo storage, where CPU speed isn't that important. Baidu is already using ARM servers for cloud storage.

Reply on news.ycombinator.com