AMD has collaborated with ARM on a chip architecture system to lower the energy envelope of server-based computing.
The success of Linux in the server world has pushed proprietary Unix and expensive server hardware based on reduced instruction set chips (Risc) to the periphery.
The vast majority of servers shipping today are powered by the ubiquitous x86 processor.
But with datacentre power requirements at a premium, chipmakers and server manufacturers are racing to make a breakthrough in balancing power and performance.
Laptop chips like Intel's Atom processor, is used to power the HP low-powered Moonshot server. This is an x86 compatible design, which means it should be capable of running 32-bit server software. AMD has taken a radical approach, partnering with smartphone processor giant ARM on its rival design.
Chipmaker ARM has been working with AMD to deliver system on a chip (SoC) processors designed for a new generation of servers.
AMD ARM-based server snapshot
- Supports OpenCompute platform, microATX motherboard using industry standard blade server housing
- ARM-based AMD A1100 series Operon CPU will ship in 2014
- Linaro tools, kernel and application support
- Server uses 64-bit ARM system on a chip with 8 cores
- 4 Mbytes L2 core and 8 Mbytes L3 cache
- UEFI standard boot with Linux supports Red Hat Fedora
- Reference board from AMD provides integrated dual 10 Gbps ports, server encryption, 128 GB DDR3 memory
- 8 lanes of PCI-Express® Gen 3 I/O
- 8 Serial ATA 3 ports
- KVM and Xen hypervisor available today
These new servers aim to tackle the biggest problem facing today's datacentres, that of power and cooling. They take a radical approach, replacing the x86 compatible processors with a 64-bit derivative of the ARM chip that runs inside mobile phones.
AMD reference server platform, claims to use 20w of power. Scaling up, 176 such servers in a two-rack enclosure based on the OpenCompute server specification would use 3.5 Kw.
“If a 100w processor is replaced by a 25w processor you can pack in more at lower power,” said Suresh Gopalakrishnan, corporate vice president and general manager of AMD server business.
By Q4 2014 AMD said that ARM powered server will start shipping. Market analysts expect that by 2019 ARM processors will account for 20% of server units shipped.
According Gopalakrishnan, almost everything in the digital enabled society ultimately ends up in somebody's datacentre. He said: "There is a lot of data in the cloud and there is a lot of parallelism, plus there is machine learning like Facebook's face recognition and even security."
All of these applications require datacentre computing. "The goal of a datacentre is to pack as much compute in a limited space. Power and space are major factors."
Datacentres are limited by throughput and bandwidth. While many datacentres are designed using one type of server to reduce operational cost, Gopalakrishnan said one size fits all does not work in a datacentre. This is because the processor core needs to wait for I/O such as to request an image from a disc drive. AMD and ARM are tackling this problem by reducing power of the processor core.
Application specific servers
"Predefined, ‘one size fits all’ server platforms are giving way to customised solutions that deliver high performance at the lowest power consumption,” said Frank Frankovsky, chairman and president of the Open Compute Project. “AMD’s contribution to the Open Compute Project expands a growing portfolio of OCP designs that enable utilisation and efficiency gains in datacentre operations."
More articles on hyperscale servers
Workloads like web services, front end and caching, storage, messaging and big data are suitable for the ARM chip, according to AMD’s Gopalakrishnan.
Moving from general purpose to application –specific servers, could revolutionise the return on investment in datacentres according to AMD and ARM. Unless server power is lowered, datacentres will be unable to cope with the explosion of data that is driving the modern world.
Rather than building monolithic IT stacks, ARM-based servers are designed to run in highly parallel hyperscale configurations, rather like the way Facebook and the search engines operate.
Ian Drew, chief marketing officer, ARM said that by 2017, every month 7.5 exabytes of data will be generated just from phones, an amount equivalent to how many grains of sand there are on the planet. All of this data will be going into datacentres.
He said the ARM-powered datacentre would offer significantly lower power and cooling requirements. "We are seeing 50 to 80% saving in power and cooling," Drew said.
The software ecosystem
There is a software ecosystem for x86 server software. Will such a system be ready for the advent of ARM-powered servers in the datacentre? Gopalakrishnan said: “Software vendors and customers are very interested. We are taking a lot of steps to make sure code is compatible.”
Lenaro is the main initiative to drive software standardisation for ARM-based servers. Linux distributions Ubuntu and Red Hat are available for the ARM server architecture. In terms of software development tools, Lenaro supports the GCC compiler the Lamp web application stack, OpenJava, Hadoop NoSQL database and PHP scripting. According to Drew, Lenaro is compatible with existing datacentre infrastructure, and so it is possible to manage ARM and x86 servers with the same datacentre management infrastructure.
Microsoft Windows Server is the missing component in ARM's server strategy. Microsoft is a member of Open Compute, it has already developed a version of Windows, Windows RT, for ARM and there is now a standard cross-platform programming model for Windows, WinRT . So the core code for an ARM-based Windows Server kernel is likely to already exist. But Microsoft's last attempt at cross platform Windows Server ended in disappointment when in the mid 90s, it stopped development of the Mips, PowerPC and Alpha processor versions of Windows NT OS.