PragmatIC 2020 predictions

19th December 2019

PragmatIC

Anna Flockett

0 0

In August 2018, Global Foundries announced a radical restructuring in the wake of AMD’s decision to take its 7nm business to TSMC, further concentrating the industry into the “mega-fabs”. The top 5 now represent 90% of the foundry business . The primary driver for this continued concentration is the cost of building, ramping yield, and efficiently operating the latest fabs, and we can be certain that in 2020 spending by these behemoths will continue apace.

We will also continue to see the development of amazing highly integrated silicon chips that will run ever more data centres and embedded control systems with sophisticated AI to help us chose our clothes, drive our cars, analyse our symptoms and much more. Another certainty for 2020 is that these ICs will get more complex, more expensive and more time-consuming to produce.

This conundrum of cost, complexity and time-to-market is also leading to the requirement for more customisation to reduce power consumption, improve performance and ultimately to reduce costs. Building a general-purpose processor from fixed IP building blocks is set to become a thing of the past, as customisable IP becomes the norm – we even saw Arm opening up its ISA this year, while SiFive continues to drive the open-source approach to processors.

We predict that there will be increasing efforts to push this optimisation even further, leading to more bespoke processors for applications where space and power are seriously constrained, including in wearables, smart packaging, digital healthcare, and interactive games. This is particularly relevant at the extreme edge of the Internet of Things – or what is now being termed the Internet of Everything. We believe that instead of using software to perform the late customisation of such electronic systems, they will instead evolve to leverage late-stage hardware optimisation; this allows application development on a general-purpose architecture, which is then compressed into just enough gates, memory and MIPS to perform the task required in the most cost-effective and power-efficient design.

An example of such an approach is the work we have been doing with Arm in creating a machine learning (ML) engine that performs a single task such as odour sensing. These ICs can be designed and manufactured, from application to fabrication, in less than 4 weeks. Coupled with the low tapeout costs of our FlexIC Foundry, this approach will radically expand the opportunities for innovators to develop compelling solutions for exciting new applications.