Since my October interview with InsideHPC last year, several readers have asked for elaboration around the assertion that ARM CPUs could disrupt X86:
The threat is definitely there. ARM architectures have been on a steady path to disrupt x86 CPUs for over a decade. Starting at the low-end with cell phones, ARM-based CPUs have somewhat quietly moved up-market into smartphones, and now netbooks. As they continue to improve in line with Moore’s Law, they are settling into 2 GHz while remaining far lower power and lower cost than even Atom. The tipping point is likely to begin in earnest during 2010.”
Two top questions have been (1) can ARM have a place in HPC, and (2) can it really disrupt X86? The short answers: yes and yes.
As recently as 2006, the idea that cell phone CPUs might someday migrate up-market into “computers” was laughable for most people. Common remarks were that “the architectural capabilities just aren’t there,” or that “the software will never get there.” That was before it was leaked that HP, Dell, Lenovo and others were developing ARM-based netbooks. That was even before the iPhone, 250,000 iPhone apps, Android, the Kindle, iPad and other intermediary steps in ARM’s upward march.
Today ARM’s growing movement into “computer-and-computer-like” platforms is old news. However the notion of ARM in HPC, or even in the enterprise, continues to raise skeptical eyebrows. “The architectural capabilities just aren’t there.” “The software will never get there.” Mark Twain said it best in that history doesn’t necessarily repeat itself, but it sure can rhyme.
In textbook disruptive fashion, ARM CPUs were significantly lower cost and lower performance than X86 CPUs along the top mainstream basis of competition. Sure, ARM ICs were cheap and low power, but they were as slow as molasses and couldn’t boot Windows. This allowed ARM to dominate a colossal, vibrant mobile ecosystem that X86 neither played in nor particularly cared about. Design wins and licensees proliferated. Billions of chips were sold.
Today, as ARM improves its clock speed (around 2 GHz and climbing) while staying lower cost and lower power, it pushes upward into more profitable, more demanding applications. Driving this trend is nothing more than the old fashioned pursuit of profit. Plotted as a function of GHz/Watt, the trajectory is revealing.
Thus the question should not be “does ARM have a place in HPC,” but “when does ARM have a place in HPC,” and “how should Intel respond?”
As the battle heats up, ARM will likely have the most appeal with customers who value low power and low cost, as opposed to raw horsepower. As such, niche enterprise and custom HPC systems are likely to provide some early beachheads. Meanwhile HPC is nothing if not keenly aware of the need for lower power and lower cost silicon – a pain point that is only likely to intensify.
For example, in this August’s issue of The Exascale Report, Mike Bernhard published a telling article (Exascale: The Beginning of the Great HPC Disruption) that quoted Sudip Dosanjh, Senior Manager at Sandia National Laboratories, and Co-Director of both the Institute for Advanced Architectures and Algorithms (IAA) and the Alliance for Computing at the Extreme Scale (ACES). Dosanjh remarked that, given the power and cooling barriers to exascale computing, “what we’re talking about here is one teraFLOPS for about 20 watts – about what it takes to run a light bulb… As a disruptive force, this would be a ‘game changer’”.
With approximately three or four ARM cores selling for every X86, ARM also benefits from attractive economies of scale. Bernhard’s article goes on to quote Sumit Gupta, Senior Marketing Manager of NVIDIA’s Tesla High Performance Graphics Processing Unit Computing Group, who commented:
“The only way to make exascale accessible to many people – is to enable the consumer market to drive the HPC processor. In other words, any government can always pay to go build one exascale supercomputer, but if you really want to make it affordable and practical for many people, research organizations, academia, government labs, industry, to have access to exascale systems, you need the consumer market to drive the HPC processor…. Let me break that down… One of the trends obvious to most people today is that the X86 CPU now dominates the supercomputing TOP 500, and the reason that happened is the economics of the X86 architecture is actually being driven by the consumer market. My laptop, your laptop, everyone’s has an X86 CPU in it.”
If consumer volume is a virtue of X86, it is even more so with ARM.
ARM, its licensees and ecosystems share a mutual incentive to move up-market into more demanding (and more profitable) computing markets over time. Today the battlefront is around mobile devices. The next frontier will be the enterprise, followed by HPC. As this trend continues, bolstered by insatiable appetites for low power and low cost processing, HPC managers would be well advised to start thinking about how to monitor and take advantage of this trend as it unfolds (rather than being caught flat-footed by dismissing it). This is not to say that ARM will dominate HPC is the near-term, nor is it to say that Intel and X86 are inevitably doomed. They are not. However, it does suggest that after three decades of relative stability, the foundations of HPC may be starting to tremor deep below its feet. There are new barbarians at the proverbial gate, and unless the right countermeasures are taken HPC may find itself squarely in the midst of an armed invasion
Thomas Thurston is President and Managing Director of Growth Science International, LLC, a research firm that predicts if businesses will survive or fail.