Summary
An application-specific integrated circuit (ASIC) is an integrated circuit customised for a particular use, rather than intended for general-purpose use as with a CPU or GPU. The two primary design methods are gate-array and full-custom design, with gate-array generally bigger, less flexible and cheaper. ASICs often ship today as system-on-chip (SoC) with integrated microprocessors, memory, and other components. ASIC design defines all the photolithographic layers of the device compared to FGPAs like FPGAs and CGRAs which contain programmable logic blocks and reconfigurable interconnects. One of the most popular examples of an ASIC is the Google TPU for running machine learning workloads.
Viability (5)
The first ASIC used gate array technology and was manufactured by Fairchild Semiconductor in 1967. The first commercial use of gate array was in the ZX81 and ZX Spectrum personal computers in the early 1980s. The field has moved from gate array and semi-custom design to full-custom design which has made the surface area smaller and improved performance. Today R&D is primary on making ASICs easier to design and cheaper to fabricate.
Drivers (5)
Moore’s Law has slowed down from the original doubling of complexity every year, then 18 months, and then every 24 months. Since 2013 and the introduction on finFET, the price per transistor has gone up. Previously developers could rely on Moore’s Law to improve raw performance faster than a product development cycle. Now developers need to consider optimising hardware for their applications. This has led to the focus on hardware accelerators, a new paradigm that pursues specialization and heterogeneity over generality and homogeneity. On the supply-side, ASIC development is easier and cheaper than ever with the development of FPGAs for prototyping and easily available IP core and open-source RISC-V. Modern CAD systems, automated layout tools, and machine learning for chip floorplanning such as deep reinforcement learning are driving down costs and improving performance.
Novelty (2)
The competitive trade-off is reconfigurability versus performance. ASICs sit at one end of a spectrum of chip designs from application-specific to general-purpose. Simply, ASICs deliver high performance but can only do one thing because they only contain the components needed for their function, so are smaller, less complex and less power hungry. CPUs and GPUs deliver lower performance but can do lots of things. In the middle are FGPAs which offer good enough performance for a few things. The downside of ASICs is that once they are made you can’t change them. So it can never run an upgraded algorithm or new specification for example. CPUs and GPUs are much cheaper than ASICs and as such the ROI will differ for every application and product.
Diffusion (4)
Of all the semiconductor technologies, ASICs and FGPAs have the easiest adoption pathways because no changes are required for fabrication. With Apple, Facebook, Google, and Amazon among the most familiar names designing their own ASICs, ASIC usage is already getting cheaper and filtering down the market. The biggest barrier is expertise designing ASICs and experience with EDA and prototyping tools. The market leaders can buy expertise as with Amazon and Annapurna Labs or hire expensive talent to build internal capacity. Because of this, we can anticipate a platform play that would drive ASIC demand as a platform aggregates workload demand and routes to the more efficient chip eliminating the need for end-user expertise.
Impact (3)
No computing technology will ever have the same impact as the transistor or more specifically MOSFET and CMOS fabrication, expect maybe Quantum Hardware. ASICs are important but are incremental innovations in the semiconductor ecosystem. They do increase performance and will catalyse AI adoption but ASICs fall in the same impact bucket as Carbon Nanotube Field-Effect Transistor as ways to extend the CMOS era. But the truly impactful technologies scoring a 4 or 5 are beyond CMOS with novel architectures or transistor types.
Sources