In April 2011 I was already arguing that AMD would eventually have to license ARM processor cores. And at the AMD Fusion developers conference in June 2011 Phil Rogers, AMD Fellow, said that Fusion processors – the ones that include CPU and GPU cores – would become CPU and GPU agnostic.
AMD as the mirror image of Intel – two hombres facing each other down main street with x86 six-guns to hand – is so entrenched in the semiconductor industry that many people couldn't quite get their heads around what Rogers was saying last year.
It didn’t seem to make many people go "woah – you mean processors from AMD with ARM cores inside?" But AMD has followed up at this year's Fusion developer summit with the announcement of the formation of the Heterogeneous System Architecture (HSA) group, with some interesting founder members and some notable absentees.
It is farsighted stuff and – in my view – the research people at AMD are rightly viewing systems from the top down and as they will be seen in three or four years' time, with hundreds and thousands of processing elements (PEs).
In other words CPU and GPU cores and other PEs are the new gates. There will be layers of software above – we hope – to provide a tractable programming model. It is the HSA group's job to make/allow that to happen.
In short, we are entering an era where programmable systems will be much closer to models of the parallel processing of the brain than to the 5F1 circuit for a Fender guitar amplifier. We are entering a neuromorphic era and leaving behind a circuit-oriented era.
It will be interesting to see whether the likes of Intel, Nvidia, Qualcomm, Xilinx and Altera can share this vision. Even if some of them DO share the vision, commercial competitiveness may persuade them NOT to join HSA on the grounds that differentiation is achieved by working separately or in a different group. Market pressures sometimes do that to technology.
Many Smartphone chips are already 'HSA' with a compliment of high speed 32 bit Application processors, multi-thread DSP, video acceleration, low speed 32 bit management processors, vector engines for the modems, acceleration blocks for cryptographic functions, HD video, audio, plus embedded power and clock management. If that isn't HSA, I am not sure what is.
I'm actually a bit surprised this hasn't happened before. AMD currently makes Intel X86 compatible CPUs. but there's no reason they shouldn't make CPUs with ARM cores. (Intel used to, in the Strongarm division they originally got from DEC sold to Marvell.) And AMD won't have any "not invented here" issues with doing so. AMD makes and sells chips, and ARM processors are poised to make a run at markets AMD is active in, so being able to offer X86 and ARM solutions is a compelling vision.
"Even if some of them DO share the vision, commercial competitiveness may persuade them NOT to join HSA on the grounds that differentiation is achieved by working separately or in a different group. Market pressures sometimes do that to technology."
This is part of the problem. The other which is inferred by the success of Intel in the PC business. In the short term it is very compelling not to compete with yourself.
David Patterson, known for his pioneering research that led to RAID, clusters and more, is part of a team at UC Berkeley that recently made its RISC-V processor architecture an open source hardware offering. We talk with Patterson and one of his colleagues behind the effort about the opportunities they see, what new kinds of designs they hope to enable and what it means for today’s commercial processor giants such as Intel, ARM and Imagination Technologies.