In the race to build better hardware for artificial intelligence, Intel is turning to an old, but unproven, type of computer processor. Neuromorphic chips, as they’re known, are modeled after the human brain, but after decades of research, have yet to show better performance in real-life applications than regular CPUs and GPUs. The veteran chipmaker wants to change that, unveiling a new neuromorphic chip this week designed specifically for R&D that the company has christened “Loihi.”
Like all neuromorphic chips, Loihi uses what researchers call “spiking neurons” as its basic computational building block. These neurons replace the traditional logic gates found in today’s silicon, and instead of processing information as binary 1s and 0s, they weight the signals they send, making their functionality more analog than binary. And, unlike CPUs, these neurons aren’t controlled by a central “clock” that regulates their calculations in a tick-tock fashion, but can instead fire as and when needed.
All these structural differences make neuromorphic chips much more efficient than today’s processors, consuming up to 1,000 times less energy. This is a huge advantage when it comes to getting AI working on devices like phones and laptops. And, the ability of neuromorphic chips to weight the signals sent inside them makes them a good match for that cornerstone of modern AI: the neural network, which processes data for everything from self-driving cars to digital assistants.
But despite the theoretical advantages of neuromorphic processors, the chips have yet to show real results outside the lab. They’ve performed well in academic and industrial research, but haven’t scale up to becoming a proper consumer product. As Steve Furber, a scientist in the field, told IEEE Spectrum earlier this year: “There currently is a lot of hype about neuromorphic computing [but] there is currently no compelling demonstration of a high-volume application where neuromorphic outperforms the alternative.” In other words, they’re just not ready for the primetime.
Intel knows this, of course, and its new Loihi chips aren’t destined for server stacks. Instead, the company will be sharing an unknown number with a few “leading university and research institutions” some time in the first half of 2018. (How many chips and who will get them are unknown.) This research will hopefully validate Intel’s designs, as well as push forward work on neuromorphic chips and AI in general.
For Intel, though, this isn’t just another avenue of research — it could be an essential counter to the rise of rivals like Nvidia and AMD. These companies have benefited from the AI boom thanks to their chip design which fits the needs of modern AI. (Just yesterday, Nvidia made a string of announcements that included new deals with Chinese tech giants for its hardware.) Other companies like Google are developing new processors for AI in the cloud, further cutting into Intel’s potential revenues.
All this has left the company fighting to hold on to its dominant position in a changing industry. So far, its response has been to buy up chip makers like Movidius, Mobileye, and Nervana, whose hardware is tailor-made for machine learning and computer vision. But with the competition closing in, it’s no surprise the company is also exploring more exotic areas of research. Chips that mimic the human brain are certainly worth a second look.