FluidCore Liquid Metal Computing Chip Released: Circuits Reconfigure in Real-Time Based on Workload

FluidCore Liquid Metal Computing Chip Released

On October 1, 2028, Caltech spinoff Fluid Computing unveiled FluidCore, the world's first liquid metal computing chip. The chip's core innovation lies in using gallium-based liquid metal alloy as programmable conductive pathways, enabling millisecond-scale dynamic reconfiguration of internal circuit topology.

Traditional chips have fixed circuit layouts determined at manufacturing time. FluidCore breaks this limitation — when executing AI inference tasks, the liquid metal automatically forms systolic array structures optimized for matrix operations; when switching to graphics rendering, circuits reconfigure within 50 milliseconds into SIMD architectures suited for parallel pixel processing.

Fluid Computing CEO and co-founder Dr. Wei Zhang demonstrated FluidCore across three different workloads at the San Francisco launch. In AI inference benchmarks, FluidCore's energy efficiency was 5.2x that of NVIDIA's H200. In graphics rendering, performance matched AMD's latest GPU at one-third the power consumption. In general computing, performance fell between the two.

The technology originates from a 2024 Caltech paper published in Nature. The team discovered that gallium-indium alloy in microscale channels can be precisely controlled via electric fields in flow direction and shape, with response times between 10 and 50 milliseconds. By designing millions of microchannels on the chip substrate with electric field control units at each junction, real-time circuit topology programming becomes possible.

FluidCore faces two major challenges. First is heat dissipation — liquid metal flow generates additional heat, with chip temperatures about 15 degrees Celsius higher than traditional chips under full load. Fluid Computing uses microfluidic cooling systems injecting coolant directly into the chip's microchannel network, though this adds packaging complexity and cost.

Second is reliability. Liquid metal's repeated flow and solidification cycles cause micro-erosion on channel walls, potentially degrading conductive pathways over time. Fluid Computing says accelerated aging tests show less than 3% performance degradation after 50,000 hours of continuous operation, but long-term real-world validation is still needed.

Initial FluidCore chips will be sold as developer kits at $4,999, including the chip, development board, and SDK. Mass production versions are expected in Q2 2029, targeting data center and edge computing equipment manufacturers.

Semiconductor analyst Patricia Moorhead comments that FluidCore represents an entirely new computing paradigm — from fixed to variable architectures. If this technology scales, it could end the separate GPU, CPU, and NPU chip landscape, covering all computing scenarios with a single chip.