Thoughts on Jem Davies leading ARM’s machine learning group
Machine learning is a new group at ARM and is not yet even acknowledged anywhere obvious on the company’s website. However, it would seem that Davies’ appointment is a tried and tested strategic move to get ARM into a business that others are pioneering and which it now wants to share in, and ultimately lead. This is what Davies did for ARM in GPUs. Prior to his leadership on GPUs ARM had been co-marketing Imagination GPUs alongside its own processor CPUs.
Over a considerable time ARM went from being an additional player in GPUs to a market leader with the combination of its Cortex processor and Mali graphics cores and the result has ultimately been the collapse of Imagination and its proposed sale to Chinese venture capital interests (see Imagination, MIPS to be sold to China-, California-connected VCs).
ARM’s move to beef up its approach to machine learning is overdue but also characteristic of the UK company’s conservative approach to technologies and markets. There has been criticism from observers that ARM has been slow to get on the machine learning bandwagon. For a couple of years startups with nothing to lose and established IP providers such as Cadence/Tensilica, Synopsys and Ceva have been introducing machine learning cores or have adapted DSPs and custom processors into vision processors and machine learning accelerators.
Meanwhile ARM hung back in 2016 (see ARM has R&D interest in neural network cores). And despite its characteristically thorough research position, the company made only tentative steps to support machine learning in software (see ARM’s soft launch for machine learning library) and on its GPUs (see ARM’s Bifrost steps up graphics, bridges to machine learning).
Next: ARM’s defense
ARM’s defence is that machine learning is a big transition for the processor industry that we are only at the start of and there is a long runway for this technology. ARM took a similar position with GPUs letting other companies be first in the market before moving decisively with the acquisition of Falanx Microsystems AS, a Norwegian developer of graphics processor cores, in June 2006.
In the case of machine learning one could have argued that ARM only needed to go at the right pace for its customer base. However, the news that Apple – a key licensee of ARM’s in the mobile space – has developed its own neural network processor (see Report: Apple working on neural processor) and numerous other developments such a $100 million Series A for machine learning IP startup Cambricon (see China chip startup nets $100 million Series A) suggests that ARM is now behind the curve on machine learning.
The idea that Cambricon has an active design-win in the Kirin 970 chip going into millions of Huawei mobile phones suggests that ARM could be two to four years behind the curve. But that also depends on whether ARM has a machine learning architecture in development and how far that has progressed. My suspicion is that there is none because there has been no machine learning group to sponsor the development; the pre-existing engineering teams have been chartered to do other things.
That said, I do not necessarily expect Davies to act quickly. If there is one lesson that could be learned from Davies’ behavior and performance when in charge of GPUs it is that both he and ARM engage for the long term, which probably means the next decade at least.
There are also arguments that while machine learning is a massive “potential” market it is also massively broad and highly fragmented in terms of the optimum network approach, architecture, datatype and data resolution. While optimizing in hardware will usually provide one or more orders of magnitude more power efficiency over GPU or FPGA approaches, there will be different optimum hardware architectures for almost every application. Some will be recurrent neural networks, some convolutional neural networks, some spiking neural networks and so on.
The machine learning circuit market could remain fragmented just as the microelectromechanical systems (MEMS) market has done; it being limited by constraints of one product, one process, one package. In the case of MEMS there are one or two key components that go to very high volume and many that remain in low volume or as custom designs.
Davies will be aware that there may be some AI workloads that will never be worth optimizing in hardware and that can stay on CPU or GPU and others that will become dedicated chips, neither of which are markets that necessarily suit ARM. One high-blown technology alternative is some form of compilible processing-in-memory architecture that would allow a broad gamut of neural networking algorithms to be mapped to a memory array (see IBM uses phase-change memory for machine learning). But I don’t expect ARM to be radical. When it was a startup it could take a risk with RISC. As a mature company and world-leader it is has proved itself no risk-taker but an almost faultless executor of the “percentage shot.”
Next: Slow and steady
So what should we expect from Davies as he moves ARM’s machine learning group towards the high ground in terms of business?
Davies will be watching and waiting to see where there are markets for products that can be monetized. In most areas of semiconductors the skill is not so much in predicting that a technology will emerge from R&D but rather in getting the timing of that emergence exactly right. Sometimes there is a first-mover advantage in markets and sometimes the first penguin off the ice floe gets eaten by the killer whale.
Davies will also be watching to find business units or startups that have some sort of advantageous position in machine learning architectures and then to make deals, including acquisitions. The successful experience of building on the start provided by engineers at Falanx Microsystems will be present in Davies’ mind. Although ARM is in machine learning for the long-haul Davies will also know that kickstarting the ARM machine learning R&D pipeline with a substantial design group will save ARM three or four years on its roadmap.
But it will also important for Davies to remember that while history repeats itself it always does so in a different incarnation. The emergence of the GPU as a key intellectual property core was in the context of the electronics as a computer-driven market that was starting to go mobile. A decade further on electronics is far more prevalent and machine-to-machine as well as human-to-machine interaction and the Internet of Things are the new realities.
Next: It’s a platform thing
ARM itself has progressed over the last decade. It has become a system-level company that underpins its system-level thinking with a variety of hardware and software licensing options. ARM has also been acquired by SoftBank Group, albeit with a remit to carry on doing what it has been doing so successfully (see ARM agrees to be bought by Japan’s Softbank).
As such, Davies may expect that platform approaches to machine learning, combining hardware and software in a system or subsystem, will be the offering of choice and that getting the application programming interfaces (APIs), libraries and development frameworks right will be the keys to success. As such second guessing who has the marketing clout to establish those APIs will be important. At this stage betting against Google and Tensorflow would seem foolhardy but it must also be remembered that Tensorflow only addresses a part of the machine learning universe, so there are opportunities for other APIs, libraries and frameworks to co-exist. Could that be the point at which ARM should apply itself to leverage success?
I therefore expect Davies and ARM to ramp up its involvement with several world-class universities in the machine learning area and to become more active as a venture capital backer of machine learning startups, either directly or through SoftBank. These would, of course, be possible moneyspinners for ARM or companies ARM could have an inside track on acquiring if developments work out.
With ARM’s CPU core business coming under some pressure from companies wishing to develop their own processor cores and the emergence of the RISC-V open source hardware, adding machine learning cores to its offerings is a vitally important step for ARM.
For now ARM says society is in the early days in the adoption of machine learning. But ARM has committed to providing resources and is putting the task of developing machine learning in a pair of its safest hands. University and startup involvement, and acquisitions are likely to follow, but it will happen over the long-term; the way ARM and Davies like to play the game.
Related links and articles:
News articles:
China’s ARM twist underlines strategic nature of tech
ARM’s soft launch for machine learning library
Imagination, MIPS to be sold to China-, California-connected VCs
Report: Apple working on neural processor
China chip startup nets $100 million Series A
IBM uses phase-change memory for machine learning
China chip startup nets $100 million Series A
ARM agrees to be bought by Japan’s Softbank
ARM-SoftBank: Selling at the top or buying at the bottom?
If you enjoyed this article, you will like the following ones: don't miss them by subscribing to :
