Xilinx CTO: Adaptive computing will continue the spirit of Moore's Law

Author: Ivo Bolsens, CTO of Xilinx

The evolution of chip architectures is accelerating, so having the right tools for the right job is critical.

The world is getting smarter. From the smartphones in our pockets to today’s smart cities managing traffic and transportation systems, AI is pervasive in nearly every industry and is impacting our daily lives. AI generates massive amounts of unstructured data that must often be managed and processed in real time. The demand for hardware is skyrocketing, with increasing reliance on innovations in chip architecture to deliver the necessary performance improvements to keep pace.

The continued improvement of Moore's Law requires thousands of engineers, hundreds of companies, and tens of billions of research and development to achieve. The same is true for artificial intelligence chips. This is not a single specification that fits all applications, nor will it be a market dominated by one company or chip architecture.

So, how can your hardware adapt to the growing demands of AI processing? The answer is Domain Specific Architecture (DSA). DSA is the future of computing, where hardware is customized to run specific workloads. DSA closely matches compute, memory bandwidth, data path, and I/O to the specific needs of application workloads. This provides a new level of processing efficiency compared to general-purpose CPU and GPU architectures.

Disadvantages of ASICs

DSAs can be built using specialized silicon, but there can be drawbacks to doing so. First, a key part of what we want is the need for faster, better, cheaper, and sooner. To keep up with the pace of innovation, manufacturers are creating and delivering innovative services in less time than ever before. More specifically, less time than it takes to design and build a new ASIC-based DSA. This creates a market mismatch between the market need for accelerated innovation and the time it takes to design and build an ASIC.

Second, ASICs are hard to make. The complexity of designing advanced node ASICs increases exponentially, greatly increasing the risk. A small mistake can have a huge impact, including at least non-recurring engineering (NRE) costs. For example, at 7nm, the cost of ASICs was hundreds of millions of dollars in NRE alone, and the cost will rise further as device geometries shrink to 5nm and beyond.

Third, ASIC implementations are not future-proof. Changing industry standards or other fluctuating requirements can quickly render the device obsolete. These are just some of the disadvantages of fixed chips.

So how can the industry continue to make architectural advances and build DSAs fast enough to keep up with the pace of innovation? The solution lies in adaptive computing.

The power of adaptive computing

Adaptive computing is based on FPGA technology, which can dynamically build DSAs in chips. Therefore, adaptive computing allows DSAs to be dynamically updated as requirements change. It frees us from long ASIC design cycles and high NRE costs. It supports over-the-air (OTA) updates of not only software but also hardware, which is especially important as processing becomes more distributed. For example, the 2011 Mars Rover Curiosity and the more recently launched Perseverance both include adaptive computing technology.

Perseverance uses adaptive computing for its comprehensive visual processing. It is built using an FPGA-based platform that accelerates AI and non-AI visual tasks, including image correction, filtering, detection, and matching. The images sent back to NASA by Perseverance will be processed using adaptive computing.

If a new algorithm is invented during the eight months, or a hardware bug is discovered, Perseverance cannot be upgraded en route. In this case, adaptive computing allows hardware updates to be sent remotely via air or space. These updates can be done as quickly and easily as software updates. When deployment is remote, such remote hardware updates are not only convenient, but necessary.

Adaptive computing can be deployed from the cloud to the edge to the endpoint, bringing the latest architectural innovations to every part of an end-to-end application. This is made possible by a wide range of adaptive computing platforms - from high-capacity devices on PCIe accelerator cards in the data center to small, low-power devices suitable for the endpoint processing required by IoT devices.

Adaptive computing can be used to build optimized DSAs in all forms, from latency-sensitive applications such as autonomous driving and real-time video streaming to signal processing in 5G and data processing for unstructured databases. With today’s hardware abstraction tools, software and AI developers can now take advantage of their strengths without having to become hardware experts.

Adaptive computing accelerates your entire application

AI inference rarely exists in isolation. It is part of a larger data analysis and processing chain, often with multiple front-end and back-end stages using traditional (non-AI) implementations. The embedded AI portions of these systems can certainly benefit from AI acceleration, but so can the non-AI portions. The flexibility of adaptive computing lends itself to accelerating both AI and non-AI processing tasks. We call this whole application acceleration, and it will become increasingly important as AI permeates more and more applications.

We've discussed the power of adaptive computing, but there are more tricks.

Start a new system

Adaptive computing makes entirely new systems possible that are not possible with other technologies. Not only does it enable rapid architectural innovation, it also enables switching between architectures while the system is running. These new systems deliver significant performance gains, power reductions, and cost reductions that are not possible with fixed chips such as CPUs, GPUs, and ASICs.

For example, in a modern vehicle, there are many cameras, each monitored by software and, increasingly, AI. Processing for the front and rear cameras is typically mutually exclusive—depending on the direction of travel, only one camera can be monitored at a time. Adaptive compute allows them to share processing resources. When moving forward, only the front video stream needs to be processed, and when moving backward, only the rear video stream needs to be processed. When the car moves from “drive” to “reverse,” the hardware is reconfigured to implement the rear camera processing algorithm, which may be different from the front-facing algorithm. This provides higher overall performance, but at the same time reduces power consumption and cost. For the end user, this means more features for less money.

in conclusion

Moore's Law is dead.

However, the spirit of Moore's Law remains.

Advances in adaptive computing and architectures such as DSAs help sustain the faster, better, cheaper quest we have developed. At the same time, adaptive computing meets the market demand for "faster" by eliminating the long design cycles required to build new ASIC-based DSAs. This trend will continue to drive rapid architectural advances, thus perpetuating the spirit of Moore's Law.