backtop


Print 19 comment(s) - last by deltaend.. on Jan 31 at 1:16 AM

The new SoCs pack 4 to 8 A57 cores, powerful coprocessors

At its fifth annual Open Compute Summit in San Jose, Calif., the world's second-largest PC and server processor-maker, Advanced Micro Devices Inc. (AMD), has announced the start of a crucial transition.  It announced the sampling of the Opteron A1100 Series processor and an accompanying development kit. 
 
I. Seattle Calling
 
These are no ordinary AMD chips.  They are AMD's first chips to pack ARM Holdings Plc's (LON:ARM) processor architecture, an architecture that has dominated the mobile space.
 
This is a soft launch.  AMD will be making actual Seattle samples (the codename for the A1100 server chips) available "in the next few weeks" in March.  Actual commercial availability will arrive sometime in H2 2014, according to AMD.
 
For AMD this is a critical release in its bid to return to profitability.  AMD has set some incredibly ambitious goals for itself, announcing its intent to become not just a leader, but "the leader in ARM CPUs" and "the leader in ARM servers".  AMD evens sets a hard target for itself -- 25 percent of total server chips sales -- more than five times its current market share.

AMD Opteron A1100
AMD presents its first ARM server chip in San Jose. [Image Source: AMD]

How did AMD get here?  What is Seattle?  And what can we expect from AMD's ARM push?  Let's look into the answers.
 
II. The Underdog
 
AMD has perpetually playing underdog to Intel Corp. (INTC).  With its challenger status has come leadership struggles, layoffs, and turbulent finances
 
AMD has tried multiple tactics to give itself competitive and cost advantages.  In July 2006, it acquired graphics card maker ATI proving to be ultimately one of the best moves AMD has made in recent years, given its sustaining profits over the last few years.  Then in Oct. 2008 it announced the spinoff of its money-losing fab business, which became GlobalFoundries.  And most recently it looked to strengthen its position in the server market with the purchase of micro-server maker SeaMicro for $334M USD.
 
But for all those attempts it still posted a loss in 2013, small as it might be.  More troublingly, AMD's market share deficit has grown, not shrunk.  In 2007 -- in AMD's strongest market (servers) -- it held roughly 15 percent of the market.  By 2012 it had less than a third of that total, with 4.4 percent of the market according to market research firm Trefis.



A major problem is simply that it's very expensive to design processors, and AMD does not have the kind of research and development cash that its rival Intel has.
 
III. The Answer?
 
With the advent of the mass mobile market, AMD sees perhaps its best chance to finally escape Intel's shadow and try to return to steady profitability.  Critically, the mobile industry operates on a fundamentally different processor design model, one that spreads the costs and development.
 
The model begins with ARM Holdings Plc (LON:ARM) which designs the basic core.  Then comes the integrators like Qualcomm Inc. (QCOM) and NVIDIA Corp. (NVDA) who add on coprocessors (media, security, I/O, etc.).  Lastly graphics firms, such as Imagination Technologies Group Plc (LON:IMG) (or NVIDIA) add on value via their special graphics processors.
 
Having watched NVIDIA's success in jumping into the role as ARM processor maker, but also NVIDIA's subsequent struggles to sell its ARM processors, AMD is planning a different route.  Instead of becoming just another mobile processor maker, AMD is looking to position itself as an early server processor maker.
 
AMD has tremendous server expertise in-house and the SeaMicro acquisition further strengthened that prowess.  AMD's server processor product line has arguably been stronger for some time than the consumer line that tries to compete almost solely on pricing.  Despite Intel's process lead, AMD's Opteron server processors in recent models have been able to beat Intel's Xeon server processors in one critical metric -- heavily threaded loads.
 
Many believe that ARM cores are more conducive to massively multicore chips.  Intel's Xeon Phi (formerly: Knights Corner) -- x86 chips that currently pack up to 61 cores -- would certainly call such claims into question.
 
But much like the GlobalFoundries spinoff, the move unquestionably will help offload an expensive and complex part of the CPU design and production food chain that AMD long did itself.  NVIDIA's ARM journey shows that ARM Holdings has made it reasonably easy for even a semiconductor firm inexperienced in processor development to deploy an ARM-equipped system-on-a-chip (SoC).
 
The move could further cut costs by allowing AMD to leverage the ARM-specific process technologies found at third party fabs like Samsung Electronics Comp., Ltd. (KRX:005935) (KRX:005930) and Taiwan Semiconductor Manufacturing Comp., Ltd. (TPE:2330) (TSMC).
 
IV. The Specs
 
The A1100 Opterons include:
  • 4 or 8 core ARM Cortex™-A57 processors
  • Up to 4 MB of shared L2 and 8 MB of shared L3 cache
  • Configurable dual DDR3 or DDR4 memory channels with ECC at up to 1866 MT/second
  • Up to 4 SODIMM, UDIMM or RDIMMs
  • 8 lanes of PCI-Express® Gen 3 I/O
  • 8 Serial ATA 3 ports
  • 2 10 Gigabit Ethernet ports
  • ARM TrustZone® technology for enhanced security
  • Crypto and data compression co-processors
AMD ARM opteron
AMD's development platform supports up to 8 SATA drives, 128 GB of DDR3 DRAM, and an octacore ARM chip.

The development platform AMD is shipping to potential customers and ARM server developers includes:
  • An AMD Opteron A1100 Series processor
  • 4 Registered DIMM slots for up to 128GB of DDR3 DRAM
  • PCI Express® connectors configurable as a single x8 or dual x4 ports
  • 8 Serial-ATA connectors
  • Compatibility with standard power supplies
  • Ability to be used stand-alone or mounted in standard rack-mount chassis
  • Standard UEFI boot environment
  • Linux environment based on Fedora, which provides developers with a rich set of tools and applications
    • Standard Linux GNU tool chain, including cross-development version
    • Platform device drivers
    • Apache web server, MySQL database engine, and PHP scripting language for developing robust web serving applications
    • Java 7 and Java 8 versions to provide developers to work in a 64-bit ARM environment
AMD ARM partners
AMD has a number of partners working on the ARM server project.

AMD is targeting large virtual machine clients, such as hosting firms, academic supercomputing institutions, and internet content providers like Facebook, Inc. (FB) (an AMD client who is specific mentioned as one of AMD's partners for the project).

Source: AMD



Comments     Threshold


This article is over a month old, voting and posting comments is disabled

Sweet!
By Argon18 on 1/29/2014 1:18:15 PM , Rating: 1
This is fantastic. We already know ARM is very capable as a server. A $42 Raspberry Pi running Linux can do 1080p video playback and serve as a great living room media center. It can also do all your standard internet things, HTTP, FTP, NFS, etc. All while consuming under 2w for the entire system. Two watts.

The potential in the server space is huge. Linux and all the open source stuff that powers today's internet like Apache, Bind, OpenSSH, etc. all runs on ARM already today. The biggest of the big are exclusively on Linux in their datacenters; Google, Netflix, Amazon, Facebook, etc. Massive power & cooling cost savings could be realized with ARM servers.




RE: Sweet!
By eagle470 on 1/29/2014 1:27:42 PM , Rating: 2
We DON'T know that this can do well as a server.

How many threads can it handle at once? What kind of computation time are we talking about? What kind of load can it handle and for how long?

While this is really interesting to see, I'll be curious to see how quickly this takes off. Given that ARM would be very invested in the idea of significantly entering the server market (sorry but HP's moonshot systems don't count) I would bet that they are helping AMD out with R&D.

Just because ARM chips use less power and cooling on their own doesn't make this viable. The limit appears to be 128GB of RAM for this system. How many of these will it take to replace an x86 Intel or SPARC system with more than a terabyte of memory? Is it still cost effective from a cooling and power stand point? Is it cost effective on a system to system level comparison?

There are a lot of numbers I would need to see before this really becomes something of interest. Also the restriction to Fedora, a non-supported (you can't buy 24/7 support) OS, keeps this out of DoD hands for now, which is still one of the BIGGEST server buyers out there.


RE: Sweet!
By Argon18 on 1/29/2014 1:32:14 PM , Rating: 2
"Just because ARM chips use less power and cooling on their own doesn't make this viable. The limit appears to be 128GB of RAM for this system. How many of these will it take to replace an x86 Intel or SPARC system with more than a terabyte of memory? Is it still cost effective from a cooling and power stand point? Is it cost effective on a system to system level comparison?"

I don't think anyone is suggesting that ARM will replace the big massive monolithic systems. You're missing the point if that's what you are comparing them to. We already know today that ARM delivers more performance per watt than x86.

"There are a lot of numbers I would need to see before this really becomes something of interest. Also the restriction to Fedora, a non-supported (you can't buy 24/7 support) OS, keeps this out of DoD hands for now, which is still one of the BIGGEST server buyers out there. "

What does Fedora have to do with anything? Fedora is for hobbyists and home use. SuSE and RHEL are the two big players in the enterprise space, mainly due to their 24/7 commercial support, and the long support timeline (RHEL6 is supported until 2024). Red Hat already has internal ARM builds of RHEL.


RE: Sweet!
By mjv.theory on 1/29/2014 2:01:47 PM , Rating: 2
Fedora is basically Red Hat (rpm), besides it's only the dev environment - it's still a Linux distro. You can always repackage applications to run on Debian and Debian based systems such as Ubuntu.


RE: Sweet!
By Argon18 on 1/29/2014 2:41:29 PM , Rating: 2
Right, that's what I'm saying. Fedora may be the initial dev environment for this developer board, but nobody in any business or government organization is going to run a production system on Fedora. Ergo, it's irrelevant that Fedora is the dev environment today. Fedora is an obvious dev choice today, since it already has an official ARM build.

You can bet that if these are commercial successful, that RHEL and SuSE will follow very shortly with official ARM builds of their enterprise operating systems.


RE: Sweet!
By KurgSmash on 1/29/2014 5:47:45 PM , Rating: 2
quote:
We already know today that ARM delivers more performance per watt than x86.


We know no such thing... We know that ARM delivers better performance per watt at low (< 1W) loads. It starts to lose that battle against higher end CPUs.


RE: Sweet!
By Ammohunt on 1/29/14, Rating: -1
RE: Sweet!
By 91TTZ on 1/29/2014 4:28:44 PM , Rating: 2
The problem is that most virtualized environments are not OS agnostic. I work in a datacenter and probably 75% of our customers run Windows Server. That'll only run on x86.


RE: Sweet!
By Ammohunt on 1/29/2014 10:23:27 PM , Rating: 1
I wouldn't say most but a lot of small to medium sized business for sure. The big players such as Amazon, Google and Rackspace who run primary Linux could see huge savings running super dense arm based virtualized environments.


RE: Sweet!
By sgestwicki on 1/30/2014 2:33:10 PM , Rating: 2
My first thought is that will be a great way to reduce power consumption at huge data centers. My second thought is that socket oriented licensing will keep most companies from using these. It will be much cheaper for my company to pay out the extra in electricity for two powerful CPUs in a server than it would be to have to get twice the licensing so I can run 4 ARM processors that can save power.


RE: Sweet!
By unimatrix725 on 1/30/2014 1:26:47 PM , Rating: 2
With AMD making it, it will have to be Price Competitive, as usual. AMD should have bought Moto instead of Lenovo. Everyone forgets the "Power Alliance" from back in the day. Moto holds a large clump of RISC patents as does IBM and Apple. Why Apple has stayed in bed with Intel is beyond me, unless for the "Sugar". I could easily see PowerPC make a comeback and even MIPs. Yes I am old, Im 34! My fist computer had a ppc. This drew me to the PS3 when it was in its prime. There have been allot of advancements on this front. Granted IBM is the leading developer here. It seems everyone lost focus when carriers began subsidizing phones. Now that T-Mobile has shook everything up. Things will and are changing.


RE: Sweet!
By 91TTZ on 1/29/2014 4:39:31 PM , Rating: 2
quote:
It can also do all your standard internet things, HTTP, FTP, NFS, etc. All while consuming under 2w for the entire system. Two watts.


That's because it's a low performance, low power device. Intel makes those too.

But there isn't much demand for that in datacenters, so Intel doesn't bother going there.

Another thing that you're overlooking is the whole x86/ARM thing. You make it sound like Intel can't compete with ARM. For one, if Intel wanted to make ARM CPUs they could. And since Intel has the best manufacturing facilities in the world they can make either the fastest or lowest power ARM chips.

Another very basic concept that you're overlooking is that performance per watt changes based on speed. You cannot compare a low-performance ARM chip to a Core i7 and then try to calculate the performance per watt that way. As you crank up the frequency on a CPU, the power consumption does not increase linearly. The power consumption increases at a much faster rate than the frequency. If you want your chip to run at twice the frequency, it's going to use much more than twice the power.

Knowing this, you can see how it's foolish to compare a low-power CPU to a high speed CPU. If you were to underclock that high-speed CPU you'd see its performance per watt increasing greatly.


RE: Sweet!
By YearOfTheDingo on 1/29/2014 7:39:45 PM , Rating: 2
Intel has gone there, actually, with their Avoton server chip. It doesn't appear to me that this ARM solution offers any performance-per-wait advantage.


RE: Sweet!
By troysavary on 1/29/2014 4:39:45 PM , Rating: 2
You think streaming media in your home is any way comparable to a datacenter?


The stupid, it burns.
By KurgSmash on 1/29/2014 3:27:26 PM , Rating: 3
"AMD is targeting large virtual machine clients.."

Yes, because so many of them are virtualizing ARM loads, and ARM is so _awesome_ in the virtualization space.

Crikies, come on.




By prsmk2014 on 1/29/2014 5:58:50 PM , Rating: 2
quote:
II. The Underdog

ARM has perpetually playing underdog to Intel Corp. (INTC). With its challenger status has come leadership struggles, layoffs, and turbulent finances.


AMD is the underdog, not ARM. Please correct. Thanks!




Not a soft launch
By BRB29 on 1/30/2014 9:42:36 AM , Rating: 2
quote:
It announced the sampling of the Opteron A1100


Sampling =/= launch

Launch is not expected until 2H 14




By deltaend on 1/31/2014 1:16:47 AM , Rating: 2
ARM cpu's might have a place in a datacenter, but only if they really scale well for high loads. For most companies that have already made massive investments in x86 farms, this doesn't really attract them at all. For those companies that currently use RISC based systems might be interested for specific applications, but I hardly see this having a major impact at all.

Regarding less power, scale those ARM chips to the performance of my XEON clusters and let's see how much power that they use. Do you remember when Intel Atom server clusters were all the rage? It didn't take long for Intel to realize that the datacenter wants BOTH more power and power efficiency all at the same time, not one over the other.




Title?
By bah12 on 1/29/2014 3:12:14 PM , Rating: 1
Didn't you mean AMD soft launches ARM server chip?




"We basically took a look at this situation and said, this is bullshit." -- Newegg Chief Legal Officer Lee Cheng's take on patent troll Soverain














botimage
Copyright 2014 DailyTech LLC. - RSS Feed | Advertise | About Us | Ethics | FAQ | Terms, Conditions & Privacy Information | Kristopher Kubicki