IBM chips away at x86 by targeting the new computing

With its latest Power8 processors IBM is capitalising on big changes to the computing order

The world of microprocessors is changing. Once, Moore’s Law – named after Intel co-founder Gordon Moore’s statement that the number of transistors in an integrated circuit will double every two years – was a fairly reliable benchmark that kept personal computers and volume servers able to deal with successive generations of software. But the compute sector today is far less uniform and today’s needs extend way beyond the ability to run Windows, Office, relational databases, email and file servers.

That alteration in circumstances is leading to something of a changing of the guard among silicon designers. Where once Intel dominated on the desktop and volume server and others occupied niches, today the rise of ARM-based chips and the concomitant increase in importance of smartphones, tablets, cloud, open source software, Big Data, multimedia and other elements is creating real competition.

One function of this change is that processor designers are taking new design routes very different to the so-called x86 instruction set Intel once made all-conquering. For example, IBM’s Power chips once targeted predictable workloads on its own servers and Apple Macintosh computers – sizeable markets but tiny markets compared to those addressed by Intel. But today IBM is looking at new markets with its recently introduced Power8, notably processor-hungry cognitive computing workloads for machine learning and analytics, an area IBM has targeted as a pillar of the company’s future in pursuit. The logic is that if it can get its Power roadmap right and execute on that strategy, customers will benefit from more efficient datacentres.

“So much of the industry is committed to Moore’s law and today the number of cores on the chip doesn’t really change things,” says Dylan Boday, senior offering manager for Linux on Power at IBM.

Boday says that the key tenet of processor design has moved from increasing transistor density to accelerating to digitally-derived insight and knocking down every obstacle that stops the brains of the computer, the microprocessor, from talking to the rest of the system.


Direct access

IBM’s latest design, marketed under the broader PowerAccel umbrella for accelerating system designs via direct connections, lets Power chips talk directly to market-leading nVidia graphics processors, or GPUs, via the latter’s NVLink interface technology. The Power8 chips will go into IBM servers but also many other designs, the firm says, with the lure of very fast computing and the ability to write software more easily because the designs are unified.

“The aim is to get data from CPU to GPU as fast as possible and open up system memory,” Boday says. “You’re no longer just stuck with graphics memory. It’s a completely flat system design with simpler code programming.”

That statement surrounding simplicity for software programmers is backed up by customers such as Bronis R. de Supinski, the chief technology officer of Livermore Computing (LC) at Lawrence Livermore National Laboratory, exactly the kind of number-crunching facility that IBM hopes to help.

IBM contends that the new designs are a good fit for categories such as open source databases and data analytics frameworks that can never get enough compute capacity. Boday claims an 80 per cent better price/performance for its OpenPower Power8 systems on some open source databases over x86 equivalents, for example.


New tricks

“We’re not going to walk away from any of those [traditional] markets but this is an open, not monolithic, environment,” he adds. Today, datacentre conversations centre around density, virtualisation, containers and new approaches to managing and moving data, he added, not just scaling up more of the same x86 capabilities.

“Speed of insight is everything,” Boday says, and these designs have the added benefit of helping to combat one of the pains of modern datacentre management: “server sprawl” where more heat is dissipated and floor space is usurped by the accretion of systems to add more compute and other capacity.

Another change: where once server makers and processor designs were focused on the needs of end-user businesses, today IBM is also very open to creating custom designs for the cloud computing giants. Armed with vast budgets and the requirement to support vast numbers of customers and traffic, these owners of enormous datacentres can demand chips tweaked to suit the way their facilities and services work. Boday won’t name names but companies like Chinese internet giant Tencent are Power customers. Tencent says it ran a data-intensive workload three times faster on a large cluster of IBM OpenPower servers compared to its former x86 platform - despite using a third the number of servers.

As more of the world’s compute, commerce and communications go through these vast data platforms the platform owners will be able to place big demands on the makers of systems and the critical components that sit within those systems. Supermicro, the hugely popular original design manufacturer that has profited from the boom in mega-datacentres, is a Power customer, as are Rackspace and Google. All of these organisations stand to benefit from what should be ferocious competition among the old guard of microprocessor designers and new names taking advantage of a fast-changing and fragmented picture.


Also read:
Does Intel regret its colocation U-turn?
IBM sets Watson free