Moonshot AI Kimi K3 Designs 45nm AI Chip in 48 Hours

Moonshot AI demonstrated that its Kimi K3 large language model can autonomously design a functional AI chip in just 48 hours. This achievement highlights a shift in how hardware engineering workflows might evolve, as AI agents take on complex tasks like chip verification and optimization. For developers and engineers, this signals that AI can now handle the full lifecycle of chip design, from architecture to final validation, without human intervention at every step.

Autonomous hardware design completes full lifecycle

The chip resulting from this autonomous process operates on a 45nm process node and runs at a clock speed of 100MHz. It features a die area of 4mm² and integrates 1.46 million standard cells alongside 0.277 MB of SRAM. The core compute unit is an INT4 MAC array with fused dequantization, designed specifically for efficient token processing.

Moonshot AI completed the entire design, optimization, and verification process using only open-source EDA tools and the Nangate 45nm process library. The company stated, "K3 is based on open-source EDA tools and the Nangate 45nm process library, and independently completed the chip's construction, optimization, and verification." This self-contained workflow proves the model's ability to manage long-range agent tasks in hardware design.

In simulation tests, the chip achieved a decoding throughput of over 8,700 tokens per second. This performance metric underscores the practical utility of the AI-designed hardware for large language model inference. The demonstration confirms that Kimi K3 can generate viable hardware specifications that meet specific computational demands for AI workloads.