Summary

Reconfigurable chips (FPGAs) are a class of integrated circuits designed to be configured by a customer or a designer after manufacturing – rather than before manufacturing as with ASICs. There are two main types: field-programmable gate array (FPGA) and coarse-grained reconfigurable array (CGRA). Both types contain an array of programmable logic blocks, and a hierarchy of reconfigurable interconnects allowing blocks to be wired together. Logic blocks can be configured to act as simple logic gates or perform complex combinational functions. Modern designs often include memory elements, which may be simple flip-flops or more complete blocks of memory. CGRAs differ from FPGAs in that they trade-off some additional configurability to achieve faster performance.

Viability (5)

The first reprogrammable logic device was delivered in 1984 by Altera, in which users shone an ultra-violet lamp on the die to erase the EPROM cells that held the device configuration. The first programmable gate array arrived in 1985 from Xilinx with 64 logic blocks. Production and circuit sophistication grew throughout the 90s with usage across the economy in particular telecommunications and networking. One of the most important applications was in speeding up Bing search with a 50% increase in throughput to accelerate search ranking. In 2022, R&D continues to experiment with different architectures, mapping and compilation strategies to deliver faster performance while maintaining high levels of configurability.

Drivers (5)

Early adoption of FPGAs was from the networking and communications industry because of frequent changes in their standards that necessitated field updates/grades, inconceivable with hard-wired ASIC implementations. The property of reconfigurability has proved valuable for embedded systems across automotive, military, and defence applications. Beyond embedded systems, new demand has come from the need to accelerate deep learning applications although this demand is shared with GPUs, ASICss, and new chips specifically designed for AI from SambaNova, Cerebras, Graphcore, Groq and Mythic. Growth in the data centre, AI training and inference and embedded systems will be core demand drivers in the 2020s especially with the introduction of memory coherent interconnect with CXL making hardware acceleration easier and faster.

Novelty (2)

FPGAs compete primarily with ASICss for workloads where a CPU or GPU does not deliver the requisite performance. Relative to ASICss, FPGAs are superior in terms of flexibility or reprogrammability, ease-of-use, and time-to-market. ASICss are superior when it comes to raw performance for a particular workload, power-consumption and mass production capability. With costs, FPGAs are materially cheaper for prototyping and low volumes, at higher volumes, ASICs become cheaper.

Diffusion (5)

Of all the semiconductor technologies, reconfigurable chips have the easiest adoption pathways because no changes are required for fabrication and they are cheap to get started compared to ASICs. Customers do not need any hardware expertise as everything is software, making experimentation relatively cheap and easy. The huge market pull for deep learning accelerators is also speeding up the development of open source developer tools including frameworks, libraries, and compilers.

Impact (3) Medium certainty

Impact is incremental like ASICs and limited to extending the CMOS era. Impact primarily on extending computing use by lowering total cost, because a chip can be repurposed and do a few things at once. FPGAs and especially CGRAs, because of the improved performance, will allow the long-tail to utilise computing and AI power. But the truly impactful technologies scoring a 4 or 5 are beyond CMOS with novel architectures or transistor types.

Timing (2020-2025) High certainty

The market was worth $6 billion in 2021 growing roughly 10% to reach $25 billion by 2030. (Compared to ASICs at $20 billion in 2021, growing at 7% to reach $33 billion by 2030). The FPGA market primarily, as CGRAs are much less mature, will grow faster than ASICs as customers look towards reconfigurability to reduce overall investment costs and extend the life of a chip.