The pace of innovation continues to accelerate, yielding more new products within shorter project schedules. The need to co-optimise electronics and the system in the context of the application workload and environmental constraints is growing. It is imperative to address system security throughout the product lifecycle as well as safety where human life is dependent upon system operation.
Design of Intelligent Systems
Artificial Intelligence (AI) has been a field of study and a technology development area for decades and is the subject of science fiction as well as everyday life. While the Holy Grail of human-like reasoning may remain just out of reach, a machine’s ability to exhibit human-like decision-making is very much a part of our lives today.
Incorporating AI into electronic systems, or as we like to call it, “system intelligence”, is certainly becoming more prevalent in the electronics industry. A multitude of AI chips were announced worldwide throughout 2019, including across Europe, and we’ll see this trend continue as AI semiconductor investments yield more results.
Furthermore, the deluge of data produced by sensors and devices is expected to reach 10 Zettabytes annually by 2025, according to the IBS report, ‘AI and Its Impact on Data Centers’. Transmitting and processing this data puts enormous pressure on semiconductor communication and compute throughput and requires new cloud data centre architectures to co-locate compute with data and employ optical connections to meet the demand.
2020 will show strong momentum when it comes to the deployment of 400 Gigabit Ethernet in data centres to handle the increasing volume of data.
What is apparent is the need to design for the holistic end application. The crown jewels of the application requirements are what’s known as the workload. This workload is the use cases, but in this data-heavy world, the workload is a huge data repository, a repository that has known provenance and represents behaviour that specific learning and decisions can be based upon. This data is crucial for correctly designing the compute and communications infrastructure. Potentially more important, it trains the system for expected decision making.
The compute performance for the overall intelligent system is highly dependent on the application and workload. Designing a component in such a system requires a complete understanding of the full hardware/software stack.
2020 will bring a growing debate about edge computing system architecture, opening questions about how semiconductor and system software works together to deliver system performance.
Billion gate semiconductors
While sensors that collect data can be small, intelligent computation is enabled by some of the largest and most complex integrated circuits (ICs) ever built. These devices integrate multiple CPUs, GPUs and, increasingly, neural processing units, or NPUs. Integrating them all is an interconnect or fabric. Fuelling the data are myriad high-performance protocol connectors. Add to that system monitoring and bookkeeping items, and these ICs integrate hundreds of large IP blocks into hundreds of millions, or even billions, of logic gates.
The complexity of these devices is immense, from implementing functionality to managing operation and data flow to supporting power-mode transitions and interrupts of various kinds. Therefore, design tools must enable more to increase the abstraction of the user’s involvement and deliver a higher quality of results automatically to augment design excellence in the resulting silicon.
In 2020, the semiconductor design community will begin to broadly adopt various machine learning-enabled design technologies to handle this complexity.
Software plays a crucial role in overall system performance. Today’s electronics environment requires co-optimisation of the hardware and software for these many use cases before the design can be considered complete - not to mention that the cost of changing hardware midway through the project is simply too expensive.
While the compute performance is a key requirement, these semiconductors and software operate in the context of a physical environment. As intelligence proliferates to more devices in more market segments, designers face a variety of physical constraints and environmental conditions. Many of these constraints impose new design criteria on electronic signal transmission, reliability, aging of electronics, thermal robustness and physical encasement dimensions. Therefore, engineers increasingly need to design, analyse and co-optimise electronics and software along with electro-magnetic, thermal and other multi-physics aspects of the system.
Semiconductor power consumption design has been pursued for several decades, primarily to keep the chips from melting during peak operating conditions or avoiding injury to the user. With the compressed form factor of many of today’s smart electronics, thermal analysis and co-optimisation of the semiconductor in the package, board and encasement environment is necessary.
Network communication speed must be high enough to support the stable operation of many types of intelligent systems. 5G promises high speed for traditional download-based use cases. But even 5G speeds may not support the kind of sampling and control latency necessary to support stable system function. And many denial-of-service (DoS) and network security questions must be addressed to realise these intelligent systems when based on 5G networks.
2020 will bring broader deployment of 5G devices and systems. More AI, data centre and communication semiconductors will be delivered as part of a solution that includes software and other system components.
Intelligence in electronic systems requires designing security in through the entire solution stack and operations. Every operation of the device, communication, data collection and management, and system operation must be protected and robust.
This begins with hardware functionality security. Semiconductor architectures must provide safe logic functionality, which includes software execution. It must protect the cache memory from inappropriate access or changes and provide a secure boot environment to (re)start a device. This includes shielding electro-magnetic data that may expose crucial data or encryption keys.
Security extends to the device as it communicates with its environment in a robust and defensible way. This includes data encryption, use of technologies such as blockchain to enforce distribution decision processes and intrusion detection.
Secure operation builds on secure communication to support over-the-air software updates. Software release and installation, configuration setting and encryption keys must all be communicated securely. The system must be monitored for correct functionality, performance of the system, fault detection and diagnosis, and fault prediction.
In 2020, expect security features to become mandatory and a must-have in semiconductor devices for more secure electronics. The promise of AI systems is compelling when assessing its convenience, capabilities and business opportunities. Designing intelligent electronics and systems must address many new challenges to realise the exponential possibilities, and semiconductor design is at the heart of these new, pervasive intelligence innovations.
Michal Siwinski, Corporate VP, Marketing Cadence