1-Bit Neural Networks & Photonic Chip Inference
~ the next wave of device-intelligence is a beam of light ~
TL;DR — Light moves faster than electrons, and ‘photonic’ chips use light for computation. They’ve been held-back by the engineering-requirements of a calculation that is used throughout Ai-tasks: Matrix Multiply. Now, with “1-bit” neural networks, those light-based chips can AVOID the over-complicated Matrix Multiply, to just ADD their signals, instead. That will allow a photonic device to *run* an existing, pre-trained Ai network as a hard-wired chip, tiny and energy-efficient. Those photonic chips will spit-out answers hundreds of times faster than electron-based circuits, for immediate reaction-time in real-world robotics and sensors. THAT changes multiple INDUSTRIES’ game.
Where Light Enters
Photonics have been in research-mode for a couple decades now, forever promising to replace our traditional computer chips, because they operate at hundreds of times the speed of our comparatively clunky silicon. Yet, that promise has been stalled, due to the complexity of building microscopic channels for that light to echo-along. Those channels needed to perform a variety of contortions, to simulate computations — multiplying numbers is tricky, when those numbers are just a flash of light in a tunnel. So, we are still waiting for photonic chips to leave the laboratory…
And now, new research into Neural Networks (the algorithms behind all our modern Ai) shows that the Ai functions fine when queried, even though their internal ‘values’ used in computation are ONLY +1, 0, or -1. They call it “The Era of 1-bit LLMs”. That was surprising!
[[ Unfortunately, *training* the Ai-model will still require huge server-centers with high-precision internal states, to compute everything exactly. The “1-bit” networks are ONLY for when you are performing ‘inference’ — that is, when you are using an already-trained-and-tested Ai to answer each specific problem. The everyday user. That’s still the lion’s share of ‘when we need compute-power for Ai’, though! ]]
So: by using this new “1-bit” method for running Ai, the internal computations are simplified from hundreds of billions of multiplication-steps, down to tens of billions of ADDITIONS. That means: we can now build photonic chips which don’t NEED to multiply! The addition of light is much simpler than multiplication, allowing addition-only photonic chips to run Ai-applications hundreds of times faster than silicon chips of today.
The Game Changed, again…
The reason autonomous vehicles and industrial robotics have stalled-out instead of taking-over the world… is their ‘Lag at Inference’. Essentially, the tiny delays of each step of computation add-up, such that the limited on-board computer-capacity cannot respond fast-enough to be useful. Bots & Autos keep bumping-into things. That same delay-during-computation is what caps performance of the US military’s autonomous fighter jets, too. Photonic chips, in contrast, are hundreds of times faster.
By building photonic chips which are *hard-wired* for specific, narrow sets of tasks, then each chip can run in tandem for a larger system with blisteringly-fast collective reaction-time. Those hard-wired photonic-chips are *unable* to change their optical wiring, but they don’t need to — they are so much more compact and energy-efficient, as a result of being engineered for Addition-ONLY, that they allow you to stack dozens of these chips together, each chip for a different Ai-specialist. (And those hard-wired specialists will never be able to modify their own internal wiring, so no Robot Uprising. whew!)
Unitree’s G1 robot, combined with photonic chips for fast reaction-times at low power, mean that you can support a large brain locally, safely, on a physically fast machine. Unitree’s G1 costs $16,000; across a 4-year recovery of the investment, that robot would need to earn you $4,000 per year. There are more than 8,000 useable hours each year — so capital costs are only 50 CENTS per HOUR! Give the G1 a photonic brain, and we can have them mop behind large equipment by hand or re-stock shelves, deliver packages, profitably.
