9MARCH - APRIL 2024just run faster every year. That was until 2005 when CPUs could not be clocked any faster due to heat. This forced AI to transition from CPUs to GPUs. GPUs are similar to CPUs, except they are massively parallel loading thousands of numbers from external memory at a time. This is powerful but not efficient since each trip to external memory requires a tremendous amount of energy. This newfound power of the GPU enabled a whole new class of ML models called Convolutional Neural Networks (CNNs).As these CNNs became popular, AI started transitioning from GPUs to ASICs. An explosion of chips started hitting the market specifically designed for CNNs by hardcoding the dataflow to keep the data on the chip, thus avoiding expensive trips to external memory. These ASICs became increasingly more powerful and efficient at the expense of flexibility. However, no matter how good they became, they were always one research paper away from becoming obsolete. For example, a paper in 2017 introduced a new ML model called Transformers. Now, Transformers are beginning to replace CNNs, rendering any ASICs optimized for CNNs obsolete.Here lies the challenge we have of putting AI chips into our cars. Three timelines that have to synchronize: software (SW), hardware (HW), and production. On the SW timeline, we have seen how AI is constantly evolving. Yesterday, it was about CNN. Today, it is all about transformers. Tomorrow, no one really knows. On the HW timeline, we can only benchmark the hardware that we can get our hands on, whereas the actual chip of interest may still be in the design phase. On the production timeline, we have to anticipate the SW and HW to sync up with the start of production (SOP) of the vehicle.To meet these challenges, we need the next-generation HW to be flexible as a CPU to adapt to any future AI workloads, as powerful as a GPU to handle increasingly larger models, and as efficient as an ASIC to operate within the power budgets of our cars. If it was the heat that transitioned AI from CPUs to GPUs and memory issues that transitioned AI from GPUs to ASICS, we wanted to explore the current trends that might take us to our goal of flexible, powerful, and efficient HW. We found three:· Trend 1- Need more flexible and efficient dataflow. Astechnology shrinks, the problem becomes the wires. They are getting too short. Data can only travel short distances per clock cycle, taking dozens of clock cycles just to traverse across the chip's surface. Next-generation HW will need to devise and create ways of masking these limitations and do it in such a way as to cater to the SW programmer. Imagine a generic dataflow design where a chip is organized as a large grid of small tiles where each tile can send data to its neighbors over thousands of these small wires in a manner that is fully programmable.· Trend 2- Need more efficient and flexible computing. All AIboils down to an assembly line of matrix multiplications. All AI hardware optimizes this. Then came Transformers and changed the game. Now, most of the matrix values are zero, with most of the remaining values are close to zero. Next-generation HW will need to be capable of automatically detecting and removing these zeros automatically, applying the minimum amount of precision on the fly as the computation progresses. · Trend 3- Once trends 1 and 2 are decided, the HW is decided.Now, it is all about the SW. With programmable data flow and powerful and efficient computing, it is now possible to write bare Meta code that can literally reprogram the chip to adapt to future models even via OTA (over-the-air) updates.The automotive and semiconductor industries can collaborate to create a seamless integration between software and hardware, resulting in optimal performance and efficiency of AI systems. This approach can identify and address any potential software/hardware mismatches early on in the development process, saving time and resources while leading to better-performing and more reliable AI hardware. By promoting collaboration, our sectors can expedite the rollout of advanced technology, paving the way for a stronger future for everyone. The automotive and semiconductor industries can collaborate to create a seamless integration between software and hardware, resulting in optimal performance and efficiency of AI systems
< Page 8 | Page 10 >