Semiconductor scaling has fundamentally changed
For about fifty years, IC designers have been relying on different types of semiconductor scaling to achieve gains in performance. Best known is Moore’s Law which predicted that the number of transistors in a given silicon area and clock frequency would double every two years. This was combined with Dennard scaling which predicted that with silicon geometries and supply voltages shrinking, the power density would remain the same from generation to generation, meaning that power would remain proportional to silicon area. Combining these effects, the industry became used to processor performance per watt doubling approximately every 18 months. With successively smaller geometries, designers could use similar processor architectures but rely on more transistors and higher clock frequencies to deliver improved performance.
Since about 2005, we have seen the breakdown of these predictions. Firstly, Dennard scaling ended with leakage current rather than transistor switching being the dominant component of chip power consumption. Increased power consumption means that a chip is at the risk of thermal runaway. This has also led to maximum clock frequencies levelling out over the last decade.
Secondly, the improvements in transistor density have fallen short of Moore’s Law. It has been estimated that by 2019, actual improvements were 15× lower than predicted by Moore in 1975. Additionally, Moore predicted that improvements in transistor density would be accompanied by the same cost. This part of his prediction has been contradicted by the exponential increases in building wafer fabs for newer geometries. It has been estimated that only Intel, Samsung, and TSMC can afford to manufacture in the next generation of process nodes.
Change in design is inevitable
With the old certainties of scaling silicon geometries gone forever, the industry is already changing. As shown in the chart above, the number of cores has been increasing and complex SoCs, such as mobile phone processors, will combine application processors, GPUs, DSPs, and microcontrollers in different subsystems.
However, in a post-Dennard, post-Moore world, further processor specialization will be needed to achieve performance improvements. Emerging applications such as artificial intelligence are demanding heavy computational performance that cannot be met by conventional architectures. The good news is that for a fixed task or limited range of tasks, energy scaling works better than for a wide range of tasks. This inevitably leads to creating special purpose, domain-specific accelerators.
This is a great opportunity for the industry.
What is a domain-specific accelerator?
A domain-specific accelerator (DSA) is a processor or set of processors that are optimized to perform a narrow range of computations. They are tailored to meet the needs of the algorithms required for their domain. For example, for audio processing, a processor might have a set of instructions to optimally implement algorithms for echo-cancelling. In another example, an AI accelerator might have an array of elements including multiply-accumulate functionality in order to efficiently undertake matrix operations.
Accelerators should also match their wordlength to the needs of their domain. The optimal wordlength might not match common ones (like 32-bits or 64-bits) encountered with general-purpose cores. Commonly used formats, such as IEEE 754 which is widely used, may be overkill in a domain-specific accelerator.
Also, accelerators can vary considerably in their specialization. While some domain-specific cores may be similar to or derived from an existing embedded core, others might have limited programmability and seem closer to hardwired logic. More specialized cores will be more efficient in terms of silicon area and power consumption.
With many and varied DSAs, the challenge will be how to define them efficiently and cost-effectively.