ACAPs are highly-integrated, heterogeneous multi-core computing platforms that significantly extend the capabilities of FPGAs and operate more efficiently than others. They are the next generation of even more powerful platforms that are already entering the market. Today’s microprocessors are so powerful and inexpensive that almost any device can be equipped with them. At present, the constant, ongoing development of these components is even enabling edge devices to be kitted out with AI.
Embedded systems and Edge Computing can only be realised for one reason. Because, microprocessors with increasing amounts of computing power are currently available with ever smaller price tags.
Moore’s law has held true for over 50 years now. It describes how the performance of computer and memory chips should double every 12 to 24 months. Accordingly, the first 4-bit processors launched onto the market in the late 1960s. They had 2,250 transistors “on board” and a clock frequency of 740 kilohertz. At regular intervals since then, processors have been launched onto the market. That enable a doubling of the quantity of processed information in the same clock period. First 8-bit, then 16-bit and then 32-bit.
At the turn of the millennium, 64-bit architecture also became available for computers. The most powerful processors currently available feature around 20 billion (!) transistors and clock frequencies of more than 4.5 gigahertz. This development was accompanied by a drastic reduction in cost. In 1961, there was an astonishing price of USD 145.5 billion (adjusted for inflation) to be paid for a GFlop. Which corresponds to one billion computing operations per second. Whereas the price today is a matter of mere cents. This means that even inexpensive, mass-produced products can be easily equipped with chips today. And that sufficient computing power is available for Edge Computing.
Lower energy consumption with ever higher performance
Nonetheless, Edge Computing is not just concerned with chip performance, but also with energy efficiency. Needless of say, many smart devices are battery-powered. ASICs (application-specific integrated circuits) offer the highest efficiency. However, these cannot be re-configured if and when requirements change.
That is why so-called FPGAs are increasingly replacing multi-purpose processors for Edge Computing. These field-programmable gate arrays are integrated circuits whose logic can be reconfigured after production. Unlike processors, FPGAs are capable of parallel data processing with their multiple, programmable logic blocks. In this case, every single processing task is assigned to a dedicated area on the chip. Which enables it to be autonomously executed. In doing so, they consume considerably less power than CPUs. As such, FPGAs combine the flexibility and programmability of software running on a multi-purpose processor. And with the speed and energy efficiency of an ASIC.
Their ability to work through many tasks at once makes FPGAs seem pre-destined for AI applications. Voice control, image processing and augmented reality are just a few examples of todays AI applications. They all require high computing power and low electricity consumption. Not to mention low latency to make the experience seem responsive and natural.
That is why the trend is heading toward transposing a growing number of AI applications from the cloud to Edge Computing. GPUs, which are frequently used in data centres thanks to their parallel-data-processing abilities, require too much power for Edge Computing. Analysts from McKinsey anticipate that the market for AI hardware in edge applications will grow. From around USD 100 million in 2017 to USD 5.5 billion in 2025. At the same time, the major chip manufacturers who currently dominate the market for cloud solutions are increasingly facing stiff competition from other established companies.
ACAPs as further development of FPGAs
ACAPs are on the horizon as potential successors to FPGAs. These adaptive compute acceleration platforms are highly integrated, heterogeneous, multi-core computing platforms that considerably expand the possibilities of FPGAs. And they are substantially faster and more energy-efficient than CPU- and GPU-based platforms. ACAPs can be modified at a hardware level and adapted to a broad spectrum of applications and computational loads. Including dynamically during operation.
Although FPGAs can only be equipped with a single logic circuit, an ACAP is suitable for accelerating a wide range of applications. Including those from the domain of AI. The first ACAP chips from Xilinx were shipped in summer 2019. They are based on 7-nanometer process technology and boast over 50 billion transistors.