Facebook's chief AI researcher, Yann LeCun, said the company is working on creating its own custom AI silicon, aiming to create much more efficient methods of processing neural networks in hardware and boost performance.
The exact details of what Facebook is doing remain unclear, although Intel announced a partnership with the company at CES this year.
Fortune, however, learned some of the themes included in LeCun’s presentation. Some of them are the expansion of AI’s role from language translation to content policing, the goal of creating smarter devices that can tell weeds from roses, for example, and computers acquiring what we usually call “common sense.”
Bloomberg, however, presents things from a different perspective. According to its reports, LeCun is focused on creating chips that don’t need to break data sets into small batches for processing, but instead work with larger amounts of information. This seems to coincide with the goal of teaching an AI-powered lawn mower to differentiate between weeds and roses. If you want to mow the lawn in a specific area, you don’t need to teach the device how to differentiate between what to cut and what not to cut, but rather teach it to avoid specific plants that you don’t want cut. The literal definition of a weed is “a wild plant that grows where we don’t want it.” A lawnmower that can target weeds but avoid roses is a lawnmower that understands which plants are desirable in a given geographic context. This is a task at which even humans can fail.
In general, the development of AI silicon – pursued by both large and small companies – is part of an effort to promote specialization.
In the early days of computing as a science, specialized architectures were simply called “architectures,” because each computer had its own operating systems, software libraries, and compatible hardware peripherals. Over time, manufacturers began to emphasize compatibility between hardware, software, and peripherals. Even in the 1980s, it was common for third-party companies to design FPUs that were compatible with Intel’s desktop, for example.
The problem with specialized microprocessor architectures, historically speaking, is that even if you had an idea for a particularly clever way to execute a particular type of instruction, the speed of general-purpose computation was increasing at such a rate that it ate away the market advantage before the product was even built. Imagine starting a company in 1990 with a chip 5 times faster than Intel's. In 1990, the fastest CPU from Intel was the 33MHz 486DX. If it took you three years to get your product to market, you would be facing the 66MHz Pentium, a CPU more than twice as fast as your original benchmark. If it took you four years, you would be facing the 100MHz Pentium. In the meantime, Intel was benefiting from economies of scale.
This unrivaled economy of scale and rapid advances in computing explain why general-purpose programs took over the market and continue to dominate to this day. GPUs are the main exception to this trend. The reason they are an exception is that the nature of graphics workloads is so different from general-purpose workloads that you could never create a GPU that could handle a serial CPU or vice versa. This is what Sony's Cell Broadband Processor attempted to achieve to some extent.
However, CPU performance scaling has stalled since Sandy Bridge. This, more than anything else, explains why Google, Facebook, and other companies are seriously considering their own architectures for specific workloads.
GPUs are expected to power the AI and ML revolution in the future. This is undoubtedly thanks to Nvidia, which currently dominates the market for these products. However, architectures related to this area, such as Google's TPU, are not going away.
Intel is already trying to address these concerns. Many of the company's acquisitions in recent years have been in the AI market, including Altera and Movidius. AMD has been focused primarily on regaining market share — its 7nm GPUs are theoretically capable of running AI and ML workloads, but Nvidia dominates this space.
Facebook’s goal of improving the use of AI and expanding the types of problems it can solve ties in with research we’re seeing from other companies. Specifically, it poses a significant threat to profits in the x86 CPU market, not because CPUs will be replaced — you’ll always need a general-purpose machine, whether it’s ARM or x86 — but because the high-margin markets where they currently sell CPUs could fill their needs with other products.
