不到 10 美元算力自动设计出四子棋专用芯片
作者提出将任意机器学习任务的数据集直接转化为专用芯片的设想,并给出一个端到端自动学习得到的单周期四子棋芯片实例。该芯片仅用不到 10 美元算力从模拟对局数据中训练生成,运行频率约 25 MHz,每步能耗仅数百皮焦,速度与 CPU 引擎相当但能耗低两到三个数量级。
报道正文
Nov. 16th, 2026
What if you could take a dataset of (input, output) pairs collected for a task of interest (image recognition, next-word prediction etc.), and turn that into a custom chip designed specifically for this task? In other words, learning an AI model where the model is the silicon? I believe this extreme specialization of hardware is where we could soon find ourselves, assuming the cost of making such chips can be dramatically reduced. This hinges on two prerequisites:
- cheaper manufacturing, from more fabs and affordable small production runs, so the world's insatiable demand for compute can be met
- making the design of custom silicon extremely cheap
Think about it: if you could cheaply print custom ASICs for every non-trivial piece of computation running over and over, why wouldn't you? Soon, AI building AI will do that anyway.
While I have no idea how to build semiconductor factories, let me convince you that automating the design of chips for arbitrary ML tasks is closer than you might think. Here is a single-cycle chip that plays Connect-4, autonomously learned end-to-end for less than $10 of compute, just from a dataset of simulated games. It has about 14k logic gates. On a two-decade-old 130 nm manufacturing process, it measures 0.8 × 0.8 mm, runs at 25 MHz and consumes a few hundred picojoules per move: about as fast as a Connect-4 engine running on a modern CPU, while using two to three orders of magnitude less energy. Try playing against the chip and see the individual wires light up within a clock cycle as the signal propagates through it: