Startup demo turns Connect-4 training data directly into a custom logic chip
A developer built a single-cycle chip that plays Connect-4, generated end-to-end from a dataset of simulated games for under $10 of compute. The design uses about 14,000 logic gates, fits in a 0.8 by 0.8 mm die on an older 130nm process, runs at 25 MHz, and draws only a few hundred picojoules per move. An interactive demo lets visitors play against the chip and watch individual wires activate as signals move through the circuit in a single clock cycle.
GoKawiil's interpretation of the reporting above, not reported fact.
The project suggests that machine-learned models could eventually be compiled directly into silicon rather than run on general-purpose processors, which could yield dramatic energy savings for narrow, repetitive tasks, as the creator claims two to three orders of magnitude lower power than a CPU running an equivalent engine. This remains dependent on cheaper chip fabrication becoming widely available, since custom ASIC production runs are currently expensive, and the Connect-4 example is a proof of concept rather than a demonstration on complex real-world AI workloads like image recognition or language modeling.
- A Connect-4-playing chip was trained end-to-end from game data for under $10 of compute.
- The resulting design has ~14k logic gates, runs at 25 MHz, and uses far less energy than a CPU equivalent.
- The approach hints at a future where ML models are compiled into dedicated silicon, pending cheaper custom chip manufacturing.
Digilent Nexys A7 FPGA Board — If learning how custom logic chips like the Connect-4 ASIC come together excites you, an FPGA board lets you prototype your own digital circuits and gate-level designs at home. The Nexys A7 is a popular choice for hobbyists and students experimenting with hardware description languages and custom logic. It's a great hands-on way to explore the same gate-level thinking behind the article's silicon Connect-4 player.
See Digilent Nexys A7 FPGA Board on Amazon → Affiliate link — we may earn a commission on purchases, at no extra cost to you. Product picked by AI based on this article; it is not a tested recommendation.Source: benoitrostykus.com, 2026-11-16
Published there as: “From data to silicon: play Connect-4 against a chip”
Read the original report → The summary and analysis above are GoKawiil's own, written from reporting by the source above. Facts and quotes belong to the original publisher.