Accelerating the AI chip design process
Artificial Intelligence (AI) is a key technology in both the hardware and software world, enabling new approaches to existing problems and expanding the realm of what is possible in a way that technology has never been able to do before. To support this technological revolution, a new generation of custom AI accelerators must be designed to provide computing performance and power efficiency for every AI application, big and small. At the same time, the High-Level Synthesis (HLS) tools that are used to design these accelerators may themselves benefit from what AI has to offer, allowing for potentially greater optimization than current heuristic models allow (to learn more about the relationship between HLS and AI accelerators, check out this blog). Further development of HLS could also allow for more integration with mechanical design tools, enabling tighter integration between hardware, electronics and software. Joining me in a recent podcast discussing these topics is Russell Klein, program director at Siemens EDA and a member of the Catapult HLS team. Check out the full podcast, read the transcript, or keep reading for a summary of our conversation.
What HLS tools, like Catapult, do isn’t magic, Russ explains; nothing the tool does couldn’t be done by an engineer. However, Catapult reaches the end result much faster, achieving an optimized design in a fraction of the time even a team of engineers could. This makes the whole process fast enough to fit within the constraints of a project’s timeframe and budget. To realize these gains, Catapult relies on a huge number of heuristics, built up over many years of design and optimization work, and customizable by designers working on a new chip.
Heuristics are, however, not inherently smart and rely on results from hundreds or thousands of previous designs to decide how a chip should function or what type of structure to include for a given task. For example, a value of a certain size may be stored in a register or memory, with each having various pros and cons. Based on historical data, the tool will decide by saying if the value is greater than X bits than put it in a register; otherwise put it in memory. While this approach does encapsulate a lot of general knowledge, it fails to account for the specific need of a particular system where, for example, the increased performance of registers may be worth using even for large values despite the greater power and area requirements because the performance is far more critical. For cases like this, AI may provide a solution.
AI offers the ability to consider not just the optimization of a single element but the impact on the performance of the system as a whole and could potentially assist or even replace existing heuristic models. Russ also highlights that AI is going to be able to start making good decisions on how to partition between hardware and software – deciding which functions should be implemented in hardware vs. just running them in software. Russ gives the example of a linear rectifier used as the activation function of a neural network, a very simple function that would normally be implemented in software but doing so creates a significant communication path into and out of the accelerator, potentially inducing a substantial bottleneck. Heuristics alone would be unable to consider the benefits of implementing a simple function like that in hardware where AI would.
HLS, at its core, is the ability to generate the design for a physical chip from various requirements. By integrating this capability with mechanical design software, it would be possible to include a physical mockup of electronic control systems based on mechanical requirements directly in the design process. Chip designs created in this way could not only help the mechanical design and EDA spaces become more integrated but also provide a place for chip designs to start, shortening the design process for next-generation AI-integrated devices.
Designing chips is a complex task, perhaps one of the most complex in the world. With an increasing demand for AI-driven smart products, the need to create faster AI accelerators that are smaller and consume less power will only continue to grow. HLS is a key tool in this process, allowing for optimization with speed humans could never match while providing a path for greater integration between the mechanical and electronics domain. In the not-too-distant future, AI itself may become a key part of HLS tools as well, offering a more complete and intelligent approach to optimization that will allow for even more powerful AI accelerators to be created in a fraction of the time, driving a loop of improvement and innovation for next generation smart product design.
Siemens Digital Industries Software helps organizations of all sizes digitally transform using software, hardware and services from the Siemens Xcelerator business platform. Siemens’ software and the comprehensive digital twin enable companies to optimize their design, engineering and manufacturing processes to turn today’s ideas into the sustainable products of the future. From chips to entire systems, from product to process, across all industries. Siemens Digital Industries Software – Accelerating transformation.
