Recently, sureCore announced results from a 28nm FDSOI test chip that showed dynamic power savings exceeding 50% and average static power cuts of 20% (when compared against a number of current commercial offerings), while only incurring a 10% area penalty for its ultra-low power SRAM IP.
And while this data is easily substantiated, the sceptical industry pundits have raised questions that fall into two camps: (a) That can't be done; or (b) How did they manage that? In answer to both of these questions, here's a quick look at the history and engineering strategy that we adopted to deliver these results.
Looking back to the early days of sureCore, SRAM fascinated us because despite many process iterations, the SRAM in use today bears a striking resemblance to the SRAM architectures that existed in the '70s and '80s. We concluded that no one had really taken a "blank-sheet-of-paper" look at the architecture for over 40 years. Recognising the growing importance of power efficiency for SoCs targeting forward-looking applications such as wearables, IoT, and other mobile devices, we examined power consumption in detail, and began by investigating how we could reduce SRAM power to a level attractive to the next generation of power critical, SoC designers.
Our starting point differed significantly from the traditional approach to SRAM R&D that typically starts at the bit cell. We recognised that the basic bit cell is fixed by the foundry; it's a piece of electronics that is carefully optimised for fabrication. Modern bit cells are designed by the foundries who tend to put an emphasis on the broadest possible manufacturability drivers; yield and faster-time-to-volume as opposed to more performance-centric metrics. Their focus is on the front-end process optimisation, area and yield.
The basic rule of R&D fabless foundry engagement has been, "use the storage array – you won’t get a better packing density." Consequently, the application use model had become separated from the technology -- ‘faster or cheaper’ became the industry’s mantra instead of ‘faster and better’. This resulted in SRAM design teams focusing on how to build more sensitive read amplifiers to detect the signals, and better write amplifiers to drive the signal on to the bit cell. Not much time was spent looking at the fundamental architecture and asking: "Is this the best way?"
sureCore decided to take a more holistic view and stood back from the whole problem. We started with a clean sheet of paper and asked, "Where does the power go when you start storing data on SRAM?"
We discovered that a lot of the power is consumed hauling parasitic capacitance around. Our design strategy was therefore very simple; we developed a system architecture to optimize power while still retaining the area advantages of the standard foundry bit cell.
Simply stated, we architected the internal block architecture of SRAM by splitting the read amplifier function into a local and global read amplifier, thus dividing the capacitive load from the word-line, only driving the areas being addressed and not the whole array. This resulted in significant dynamic power savings during the read cycle. In a similar fashion, we reduced the write cycle power by a similar amount. Whilst hierarchical solutions are not new, the sureCore “secret sauce” is at circuit level developed by our engineering teams leading to not only significant power savings, but also comparable performance levels.
Our "blank sheet" approach delved deep; right down to the fundamental device physics level. Our strategic partners, Gold Standard Simulations -- recognised world leaders in modelling devices at the atomic level and experts in nano-scale process nodes, helped us to understand the behaviour and limitations of processes at nodes below 28nm at a device level and bit cell level. Combining this fundamental device understanding with excellent circuit design and system analysis skills, we’ve identified where existing SRAM solutions waste power, and architected our solution to avoid this; we deliver power savings without the added complexity of write and read-assist.
At the outset, we determined it was important that our IP be process-independent. sureCore IP is based on architecture and circuit techniques rather than a reliance on process features. The result of this is technology that can reduce power in standard bulk CMOS, but is equally applicable to newer FinFET or FD-SOI processes and across all geometries, even down to 16nm and below. We believe our approach is paying off and, because we insisted in retaining the foundry optimised bit cell, sureCore's technology can be retrofitted into existing designs enabling extended product life cycles.
This is our basic technology story... a start-up deciding to take a fresh look at an old technology and improving power performance over 50%. What do you think of this approach? Do you have application ideas for our new, ultra-low power memory? Send your thoughts to me at firstname.lastname@example.org and I'll post them for all of us to discuss.