ARM server hype ramps faster than ARM server chips
All the more time for Intel to get a leg up
The next-gen X-Gene
The next generation X-Gene will be shrunk using TSMC's 28nm process, and will have a total of 16 cores running at 3GHz. The coherent network on the chip will extend out to 64 processor sockets in a glueless fashion to a maximum of 1,024 cores in a cluster image. While that is nowhere near the scale that Calxeda is talking about with clusters based on current and future EnergyCore SoCs, it is still a large number and something that cloud providers will take a hard look at if the chip and its interconnect work as advertised.
That 1,024 core count is certainly a lot higher than Intel is promising as a single cluster image for either Xeon or Atom processors with integrated switching, since Intel has not promised any integrated switching – yet. All that it has said is that the future "Avoton" Atom S Series chips due later this year will include on-chip Ethernet controllers. Intel was hoping to build an external switch business that rivals its server and storage biz, and here are all of these ARM vendors talking about integrated switching.
Marvell: It has been more than two years since this chip maker launched its Armada XP ARMv7 derivatives aimed at servers, and the company has gotten some traction with its silicon.
Dell is not playing favorites in the early adopter phase of the ARM server market, and is shipping hyperscale servers based on Calxeda ECX-1000 processors called Zinc, which slide into its C8000 chassis, as well as the Copper server nodes based on Marvell's ARM processor and networking chips, which slide into its C5000 chassis.
The Marvell-based Dell machines use the 40-bit, quad-core Armada XP 78460 processor, based on its "Sheeva" family of SoCs. The Sheeva PJ4B cores run at a modest 1.6GHz, and each SoC can address up to 8GB of DDR3 main memory (which has ECC memory scrubbing).
There are four PCI-Express 2.0 controllers on the Armada chip, plus controllers to drive two SATA peripheral ports and four Gigabit Ethernet ports. The chip has a 4Gb/sec packet processor, perhaps useful for encrypting and decrypting data, and a controller to link to three USB 3.0 ports – all packed into a 15 watt thermal envelope. Dell chose the companion "Cheetah" Layer 2 network chip to link multiple Sheeva SoCs together.
This is the same basic Marvell chip technology that upstart server maker Codethink is using in its Baserock Slab ARM server, by the way.
Marvell got into the ARM chip racket after it bought Intel's Xscale ARM chip biz back in 2006, when Intel's top brass decided to focus solely on the x86 architecture for its processors. Notably, when the big Cortex-A50 announcement was made at ARM TechCon 2012, Marvell was not one of the licensees. But you can license the ARM specs and create a custom chip rather than license the ARM Holdings reference design, as do Apple and Nvidia, just to name two companies, so Marvell could be working on a next-gen Armada XP aimed at servers.
Nvidia: Everybody was sure during the 2010 rumor mill that graphics chip maker Nvidia was going to get into the x86 processor racket, taking on Intel in PCs and servers. They were wrong.
What Nvidia did announce in January 2011 was an effort called Project Denver, which will create custom 64-bit ARM processors that the company will embed on future "Maxwell" GPU chips, due in 2013 and offering around five times the gigaflops per watt as the current "Kepler" series of GPUs.
Nvidia's Tegra line of smartphone and tablet processors is based on ARM architecture, so the company has plenty of experience designing its own ARM variants. To a certain extent, Project Denver is not something new so much as an effort to unite two parts of Nvidia onto a single die. Nvidia has not revealed much about the Denver effort, but CEO Jen-Hsun Huang said during a conference call going over Nvidia's Q3 numbers in early November that Denver "is going great" and that it was an ARMv8 processor with "some exciting secret sauce."
AMD: If AMD has a code name for its future ARM Cortex-A57 Opteron processors, El Reg doesn't know about it – but perhaps "Baking Soda" is a good name, seeing as how AMD hopes to take the best of its Opteron processors, which were originally known as the "Hammer" family, and match it up with ARM cores to make an x86 alternative for servers.
AMD has not yet provided a roadmap for its ARM-based server chips, but said last October 2012 when the Cortex-A50s were launched by ARM Holdings that the processors would carry the Opteron brand and that the "Freedom" interconnect fabric that is at the heart of its SeaMicro microservers would be embedded in some fashion on the chips, allowing "hundreds of thousands" of CPUs to be linked into a cluster over that fabric.
These ARM-based Opterons are expected in 2014, and while AMD didn't talk about it, it's also possible for AMD to roll up ARMv8 chips to serve the tablet and PC markets, as well.
Cavium: This supplier of network processors based on the MIPS architecture is expanding out to ARM chips through an effort called Project Thunder, launched last August. The details are quite sketchy about Project Thunder, but Cavium is a licensee of the ARMv8 architecture and says that it is building a custom chip rather than licensing either of the Cortex-A50 chips as the basis of its designs.
Cavium has processors based on MIPS architectures that span from two to 48 cores, and it will be interesting to see if the company will push up the core-count envelope, push down the clock speeds, and go after parallel and media workloads much as it has with its Octeon MIPS chips. The company says it is aiming at compute and storage workloads in the data center with the Thunder chips, and will give out more details about its plans in the future.
The word on the street in the wake of Cavium's announcement last summer was to expect Thunder chips in the second half of 2014. That seems like a million miles away, but it is when many other 64-bit ARM server chips are also expected.
Samsung: In many ways, Samsung is the wild card in the ARM server-chip racket. The company has big advantages over other ARM server chip providers in that it also makes and sells its own main and flash memory as well as disk drives. The Korean giant could do some very clever integration if it licensed or developed 3D packaging technologies.
Notably, Samsung was a licensee of the Cortex-A50 series of chips, but the company has not responded to chatter about its rumored plans to enter the server fray. All the company said back in October was that it was working on "future computing platforms." No kidding.
With so many companies working on processors and interconnects, the open source Linux operating system is a natural fit for the ARM chips, since chip developers and software providers can help get Linux working on all of these different machines.
But the variety may be as much a hindrance to the adoption of ARM chips in the data center as it is a help. The chip makers are going to have to get enthusiastic and consistent support from the Linux kernel coders and distro makers to support many of these chips before they become economic successes, and that is going to take a certain amount of money as well as time.
So the ARM server chips with the most enthusiastic support in Linux may ramp up even if it is not necessarily the best technical option. For instance, an interconnect could be problematic to support with the Linux kernel and thereby limit its appeal – though we are not saying that this has or will happen.
Limiting ARM servers to Linux only is also an issue over the long haul, but is probably not a big deal in the short run. Those hyperscale cloud providers (with the exception of Microsoft) that write their own code for the most part to provide their search, social media, and other services are already running on Linux. And if ARM takes off at Microsoft's competitors and provides a substantial cost or integration advantage, then that is when we can expect Redmond to roll out a Windows Server 2013 edition for ARM servers.
Microsoft will no doubt want to be its own first customer for such a software release. ®
Sponsored: Benefits from the lessons learned in HPC