The challenge with digital computation in AI
Over the past few years, AI has exploded in use and presence, accelerated by the many benefits it provides, but it faces power and efficiency challenges because of the massive amount of computation it requires. One problem that AI faces is that digital computation has inefficiencies for many AI workloads, particularly in energy usage and parallelism. This is why CPUs are rarely used to run large-scale AI models; computations are more often carried out on GPUs and, increasingly, on NPUs. Even then, these devices are still not fully efficient for all AI tasks, and as such NPUs are becoming more important in embedded systems. Part of the reason is that biological neural processes are analog (connections grow and shrink depending on usage), and many researchers argue that representing weights as continuous physical quantities could be more efficient. An analog computer could simulate such behavior by having each “neuron” be a control element whose conductance or resistance (for example, memristors) or magnetization changes depending on signals received, and these changing properties would affect the flow of current through the AI system. A mostly analog implementation could compute many operations (such as matrix-vector products) in physical time with massive parallelism, potentially offering much higher speed and energy efficiency for certain tasks; however, it would not be truly "instant" and faces challenges like noise, limited precision, and programmability. Such analog‑oriented chips could be made small and energy‑efficient enough to embed in many devices, which is why analog approaches are of interest for future AI systems.
Startup Company Unconventional AI Rasies Funds to Create Analog Solution
Recently, a startup company called Unconventional AI raised $475 million in seed funding at a $4.5 billion valuation, backed by Andreessen Horowitz to develop an analog solution for AI workloads. The co-founder of the company, Naveen Rao, previously co-founded Nervana Systems, which was acquired by Intel (and he later worked on AI efforts there). Considering that analog computers from the past have proven to be extremely energy efficientm Naveen Rao has turned his attention to creating a new analog chip that could run AI algorithms efficiently. However, unlike analog computers used in the past, Naveen Rao is looking to create a chip that can scale with the latest technology processes to improve AI performance. But why is this important? When it comes to AI, improving accuracy generally requires more data to train from, and this leads to larger models. As such, AI engineers will deploy hardware with greater memory capabilities to handle the increased size of their AI model. However, the use of digital circuitry in the form of transistors fundamentally limits what can be done. For example, as transistors shrink to smaller process nodes, limits on current-carrying capacity and power density constrain how many cores can be integrated into a single chip. This is problematic when trying to scale AI models as digital processors are not ideal for the job. To solve these challenges, Naveen Rao is looking towards analog circuits to solve the AI challenge. Of course, analog computers are notorious for being difficult to scale and maintain precision, but their extreme energy efficiency (the human brain consumes roughly 20 watts, whereas large AI systems consume orders of magnitude more power across datacenters) makes them far more appropriate for AI processing. Unconventional AI hopes that by building their first large analog chip with TSMC over the next five years, targeting neural-network modeling workloads, they will be able to solve the problem of AI energy efficiency while creating a chip that could revolutionize the field of AI.
Is Analog Really the Answer?
Is Analog really the right answer to future AI systems? While in the real world, brains and nerves are analog, in the digital world, this may not be practical for digitally interfaced systems. Of course, analog neural networks have been demonstrated to function well, but they are not easy to program, train, or reconfigure. This is where exotic components such as memristors come into play as they can be programmed to hold a specific analog value. But what makes Unconventional AI interesting is that it has managed to create an analog AI system that can be reprogrammed and trained. If Unconventional AI can continue to improve its technology and develop more complex hardware, it could very well be the foundation for future analog AIs, and that would be a true revolution in the AI industry.
You may also like
