By RK Anand, Co-founder and CTO at Recogni.
Reducing Power Consumption is a Top Priority Across All Industries
Whether we’re talking about electric vehicles or data centers, there is consensus that power consumption is skyrocketing and will only continue to get worse, especially with the increased use of AI. There will be a strong focus, and investment, on technologies and processes that can reduce power consumption and utilize energy more efficiently. Everything from massive compute architectural innovations, high performance power-optimized silicon, to new materials and creative cooling approaches are going to be key to enable an AI-first world
Generative AI Performance Set New “Standard” for Speed
Not only is the quality of generative AI output at a relentless pace of improvement, but the speed at which the results are delivered are unprecedented. As businesses and consumers alike come to expect this kind of speed with results, this same speed of progress will be expected of other industries and technologies – like factory automation, automated farming, accelerated construction and autonomous vehicles. There will be an expectation that any ADAS or self-driving feature (regardless of SAE Level) will perform nearly instantaneously and flawlessly. This means that there will be little room for error to regain public trust in self-driving technology, especially given the recent debacle with Cruise in San Francisco. There will be no room for “hallucinating” or “phantom breaking” – the technology has to be 99.999999% perfect, which means fast, accurate and power efficient.
Self-Driving Will be 100% AI-Based
The rate of progress in AI is unrelentless and we are seeing the quality of results with generative AI, especially with larger and larger models, is getting better. There are indicators for autonomous driving that it is within the realm of possibility, in the next three to four years, that the end-to-end method for driving autonomously will be fully AI based. Right now, the front end is detecting objects, lanes, etc. and using AI to make determinations, whereas the rest of the driving path (prediction, path planning, controls, etc.) is a more traditional standard computing approach. There are now indicators that in end-to-end AI based systems – one could potentially go from sensory input, full comprehension to vehicle level control actions. The critical element here will be to validate such systems, get to the point of “explainable-AI” and guarantee safety goals. This makes the full end-to-end driving loop fully AI-based. Of course this means that workloads will increase (i.e., a large model that could typically be run in the cloud will now have to run locally in a self-driving vehicle), which means increased power consumption, and leads us back to our issue around finding solutions to reduce power consumption.
