Almost as soon as ARM-based server company Calxeda Inc. (Austin, Texas) had said it could not continue its operations in December 2013 than another ARM-based server company, Rex Computing Inc. was wowing them in the aisles at the Open Compute Project summit in San Jose, Calif.
So what is it about Rex that makes the company different and likely to do better than Calxeda. And what does it say about the evolving semiconductor industry.
Well the first thing to state is that the closure of Calxeda was one of very few set-backs for the ARM ecosystem in recent years. Secondly it does not mean that ARM-based processors will not be introduced in that space. Applied Micro Circuits Corp. (AMCC) is launching the X-Gene 64-bit processor for server applications and Advanced Micro Devices Inc. has just launched the Opteron 1100 series 64-bit ARM processor to name just two major chip companies attacking the market.
And "major" is the operative word here.
So one thing that marked Calxeda out was the fact that it was a startup fabless chip company. It choose to develop a 32-bit ARM server processor first. And it looks like that was a failure of market-entry timing in that the company failed to get enough traction at 32-bit to fund its ambitions at 64-bit where it was falling behind the larger chip companies.
And one thing that marks Rex Computing out as different – apart from its 17 years old CEO, Thomas Sohmers – is that it is not trying to develop chips. Instead it uses $99 Parallella boards from Adapteva Inc. and puts 16 together to form a computing fabric in a server blade. Those Parallella boards include the dual-core Cortex-A9 on a Xilinx FPGA and a 16-core Epiphany III 65nm processor from Adapteva. Rex Computing is putting its effort into writing the software that will allow the processor farm on that blade to get to grips with problems efficiently.
So far, so sensible, in that Calxeda managed to burn through more than $100 million designing proprietary ARM-based silicon and has failed to make the cut.
However, before you rush away with the idea of Calxeda and leading-edge chip development "bad" and Rex Computing, software development "good," one needs to look a little deeper.
For sure Rex is pushing its risk upstream in the supply chain but that doesn't mean that Rex is not still subject to that risk. Adapteva was founded in 2008 by a former Analog Devices design engineer Andreas Olofsson, Adapteva has developed a series of processors on meager resources mainly by using multiproject wafer runs. But in 2012 the compay turned to crowd-finding website Kickstarter to try and get funds for an Epiphany processor mask revision. It raised nearly $900,000 and considerable awareness for its Parallella board, but has struggled to fulfil promises to provide boards to those Kickstarter supporters. Hopefully, those were board-level teething problems, now resolved.
But the issue for Rex is that if Adapteva should fail to deliver on its roadmap to 64-bit processing then Rex will be seriously compromised in the market, just as Calxeda was. In other words its risk is almost as high as if it were a fabless chip company. It does not have the comfort of knowing its destiny is in its own hands and it still has the challenge of developing software that can get the gigaflops out of the server blade and demonstrate improved energy efficiency over the competition.
What this does show is that designing digital silicon at the leading edge has become a significant challenge for the startup company. I am sure that readers will be able to point to startups that have got 28nm silicon back from TSMC and are making good money on them. But in this case Calxeda has missed its mark and Adapteva is only just coming to the leading edge with Epiphany IV on 28nm.
Part of the equation here is that the cost of chip development is going up exponentially with each generation and much faster than the value of most of the available markets. In other words the profitability is going out of the semiconductor industry. This then increasingly favors consolidators that have economies of scale or the vertically integrated companies – such as Apple and Samsung – that can cross-subsidize chip development from equipment revenues.
Related links and articles: