Applied Micro reaps the benefits of X-Gene ARMv8 processors
4 mins read
In 2010, Applied Micro made a strategic decision; courageous or foolhardy, depending upon where you sat in the communications processor world. That decision was to sign an architectural licence for ARM's 64bit v8 processor cores and to embark on the design of a multicore device.
The company's incoming CEO at the time perceived the need to rethink the architecture of server processors in order to achieve a better balance between power and performance and to make such processors more attractive economically.
In itself, the decision to follow the ARM route was not an issue; the interesting factor was that ARM had still to finalise the v8 architecture. It was, to a certain extent, a case of the chicken arriving before the egg. Early involvement allowed Applied Micro to help ARM to complete the v8 specification and to write more than 20,000 instruction set verification tests.
Spin forward to 2014 and the fruits of Applied Micro's labours can be seen in the shape of X-Gene, although the development process passed an interim stage in which an FPGA based version of the processor was created.
The reason for the interest in the v8 architecture – not only from the supply side, but also from developers – is the need to reduce the amount of power consumed by server farms and the like. Instead of chasing raw performance, datacentre developers are now looking to maximise parameters such as performance per Watt, as well as getting the most performance for their money.
Gaurav Singh, Applied Micro's vp of engineering and product development, said: "X-Gene is an ARMv8 compatible design. It's a high performance CPU compared with existing ARM designs and we're looking for the device to provide 80 to 90% of the performance that operators would expect from a high end Intel Xeon processor."
Singh described X-Gene. "It's a four issue out of order CPU running at 2.4GHz. It has a range of high end features and includes performance optimisation for hypervisors."
Applied Micro is one of a number of companies looking to displace Intel from its position as the dominant supplier of processors to datacentres and similar demanding applications. Others lining up against Intel include Cavium and AMD, both of whom have launched or are about to launch ARMv8 based parts into the sector.
AMD announced its intention to enter the 64bit server market in October 2012. At the time, Rory Read, its president and CEO, said: "Through our collaboration with ARM, we are building on AMD's rich IP portfolio, including our deep 64bit processor knowledge and SeaMicro Freedom supercompute fabric, to offer the most flexible and complete processing solutions for the modern data centre."
But, while AMD is pursuing what it calls an 'ambidextrous' approach – building ARM based server chips alongside its existing x86 portfolio – Applied Micro is following an ARM based strategy. The first implementation of X-Gene is an eight core device with a high speed interconnect. "It features an 8Mbit L3 cache," Singh noted, "and lots of memory channels."
That last feature is something which Applied Micro believes it is necessary to address. "One of the things that's happening in workloads for the cloud," he continued, "is that a lot of the applications are memory bound. So we have integrated four DDR3 channels." But another key technology in X-Gene builds on Applied Micro's expertise in data transport. "We have a separate transport group which has been designing high speed SPIs and so on for telecoms manufacturers, amongst others, and the group is a leader in developing 100G OTN solutions. We've been able to bring this technology into X-Gene."
In its pitch for X-Gene, Applied Micro lists a number of benefits for the device, including 'enterprise class' error correcting code and reliability, availability and serviceability, or RAS. It also points to 'highly integrated mixed signal I/O' and has integrated a range of interfaces into X-Gene, underlining the importance of connectivity in its target markets.
"There's a lot of high speed I/O," Singh pointed out. "As well as four 10G Ethernet ports and a USB port, there are 18 lanes of PCIe gen 3 and six SATA ports for storage needs. We've looked to integrate all the components needed for high performance, power efficient computing and are taking that device to production."
One particular function which has been designed into the processor is the extravagantly named RoCE – remote direct memory access over converged Ethernet. This takes advantage of Applied Micro's expertise in Infiniband. Singh explained the significance. "A lot of data centre apps are affected by latency, but the current trend is for a lot of apps to talk to a lot of other apps. Most of that communication is currently enabled by TCP/IP, but that's a legacy protocol. Although it works well and underpins the Internet, it's not a protocol which is conducive to short transactions because its overhead gets in the way.
"Using RoCE, that overhead is taken away, so something that used to take 30µs can now be done in 5µs."
Although the X-Gene project has been underway for some years, the chip itself has appeared on the market almost by stealth. Production of a 40nm based version of the device started in March 2014, with a 28nm variant expected to sample in the near future.
"The Gen 2 device at 28nm is partly a process shrink," Singh said, "but we have also made some optimisations to the CPU cores, while taking advantage of the higher frequency and lower power consumption of the 28nm process." In fact, Singh said he expects an 'incremental improvement' in performance of between 15 and 20%.
Applied Micro is now working on the next generation of X-Gene. "We'll be moving beyond eight cores in the future," Singh said, "taking advantage of the GICv3 specification, which is likely to be ready for 2015 or 2016." And he said that X-Gene 3 is being designed for manufacture on a 16nm finFET process.
While data centres are an obvious target market for X-Gene, there are others. One is high performance computing and Applied Micro was a major presence at the recent International Supercomputing Conference. Another is hyperscale computing, which requires extra resources to be added as demand increases. "There's an explosion of data," Singh said, "and a lot of data processing now needs to be done simultaneously. A lot of customers are getting excited about X-Gene."
X-Gene powers server motherboard
UK developer SoftIron says it is rewriting the rules of high performance computing with the launch of a board that delivers 'significant reductions in operating costs, at around twice the performance per Watt, and significantly lower temperature operation than traditional systems'.
The SoftIron 64-0800 is the first product in a range that will include single processor servers, high density blade servers and motherboards for self assembly.
SoftIron's CEO Norman Fraser said: "Our design takes X-Gene and integrates it with an efficient board and power supply design to deliver server technology that maximises efficiency and improves the total cost of ownership."
The board also provides hardware support for standard cryptographic algorithms and protocols, improving the performance of compute intensive security tasks.