Xilinx CTO: 2020, use Kurzweil's law to accelerate return on investment

Some exciting new technologies are emerging, such as 3D integration of multiple chips: planar layout has already utilized every square millimeter to the extreme, and the next step is to tap the potential of three-dimensional space; rapid gains can be achieved by introducing new technologies such as storage-level memory and silicon photonics. When encountering the physical speed limitations of conventional I/O circuits (approximately 100Gb/s per meter), it is necessary to increase the speed of multi-chip connections and reduce I/O power consumption, which brings silicon photonics technology to a new generation of advanced ICs such as FPGAs.

This is indeed a very exciting time. 2017 Turing Award winners Stanford University Professor John Hennessy and Berkeley Professor David Patterson have both hailed a golden age of computer architecture, one of the key drivers of which has been the pursuit of optimization in specific areas. Take the Xilinx AI Engine, for example, one of the most important and powerful features of the Versal™ ACAP adaptive compute acceleration platform launched by Xilinx in October 2018. A few years ago, Xilinx would have been unlikely to launch such a platform because, at that time, performance improvements would have been more easily achieved through other means.

Today’s progress isn’t just in processing performance. Earlier in this article, we mentioned the development of silicon photonics, a technology used to increase I/O speeds. In fact, the explosion of AI workloads is one of the most powerful drivers that has led Xilinx to focus on accelerators, such as the Versal chip, and seek ways to speed up the flow of data into, between, and out of the accelerator. Versal’s programmable, high-bandwidth network-on-chip (NoC) interconnect, as well as other design features such as the tight and short connections between distributed on-chip memory and processing units, are examples of chipmakers looking for ways to further extend Moore’s Law to achieve a new generation of performance gains.

Figure 1: Optimizing internal device interconnection to improve processing speed and efficiency

AI is likely to be the primary factor influencing current and future processor architectures. It is undoubtedly the application needs of the data center that drive much of Xilinx’s work today, which revolves around the diversification of workloads.

Historically, hyperscale data centers have been huge repositories of data archives, storing videos, pictures, audio and other content, and even providing on-demand content services. As large amounts of IoT data from self-driving cars, smart factories, smart cities and smart infrastructure are connected, all of us (both individuals and businesses) have higher requirements for data processing. Enterprises need to quickly obtain much-needed insights and decision-making power from the messy and huge data to continuously improve corporate productivity, energy efficiency, public safety and security, and our living standards.

Given the diversity of workloads, data centers require a variety of different resources to effectively address them. Data center architectures are moving away from rigid, CPU-centric structures to prioritize the ability to flexibly respond to change and configurability to optimize resources such as memory and accelerators assigned to individual workloads, without a single value standard . It’s not all about Tera-OPS. As demands become more real-time, other metrics such as transfers per second and latency become more important, with self-driving cars being an obvious and important example.

Figure 2: Diversity in cloud workloads requires greater flexibility

Obviously, this area is exactly where Xilinx's unique programmable devices and expertise excel, and Xilinx is also driving adaptive computing acceleration platform solutions such as Versal ACAP to meet the above industry needs. Looking back at Xilinx's early business, it was mainly aimed at ASIC designers who were looking for faster design cycles and lower engineering costs, as well as EDA software users. From this perspective, today's Xilinx is no longer the same. Today, Xilinx's user base is expanding to computer scientists and data scientists, which makes Xilinx pay more and more attention to how to provide powerful tools so that users can fully utilize and benefit from the advantages of Xilinx's powerful programmable devices and platforms without having to understand the details of the underlying architecture. I firmly believe that the PYNQ™ - Python on Zynq program that Xilinx is currently actively promoting will be an important advancement that will make Xilinx's cutting-edge programmable architecture more popular in a wider community .

Another big industry trend is blockchain. To some, blockchain may have a bad reputation because of its association with the anarchy of cryptocurrency, but I believe that blockchain will have a wider impact than most of us realize. When ARPANET first emerged as a simple platform for distributed computing and sending emails, who could have foreseen the development of today's Internet? Through projects such as open source Hyperledger, blockchain technology may be a game changer and become the platform for establishing trust in Internet transactions.

The trusted Internet may soon become a hot topic in the industry, allowing people to prove data without having to submit more data, thereby effectively protecting privacy and ultimately solving problems such as fake news by clearly identifying the origin and source of information. The industry needs to find effective ways to build and scale blockchain applications, and technologies like Xilinx ACAP that can effectively accelerate computing, storage, and networking will undoubtedly be a major component of the solution .

Figure 3: Configurable programmable accelerators enable future trusted Internet nodes

The predictability of Moore's Law may have become too smooth and slow. In the future, we need to maximize the flexibility, agility and efficiency of Xilinx technology and extend it to some communities that may not be familiar with these technologies, because if we want to achieve the progress that everyone needs, we must absorb their wisdom. As Moore's Law becomes a thing of the past, we can see the inevitable trend of Kurzweil's law of accelerating returns more clearly .