LLM Agent Builds TurboQuant Inference Accelerator in 80 Hours
A multi-agent system called Design Conductor 2.0, powered by frontier LLMs released in April 2026, autonomously designed a TurboQuant inference accelerator named VerTQ in 80 hours. The system handles tasks 80 times larger than its predecessor, which built a 5-stage RISC-V CPU in 12 hours in December 2025. VerTQ hard-wires TurboQuant support in a 240-cycle pipeline, includes 5129 FP16/32 compute units, and was mapped to an FPGA at 125 MHz, consuming 5.7 mm² in TSMC 16FF with 8 attention pipes. The system produced four autonomous designs total, starting from the TurboQuant arXiv paper.
Key facts
- Design Conductor 2.0 built a TurboQuant inference accelerator in 80 hours
- The system is powered by frontier LLMs released in April 2026
- Prior version built a 5-stage RISC-V CPU in 12 hours in December 2025
- New system handles 80x larger tasks autonomously
- VerTQ accelerator hard-wires TurboQuant in a 240-cycle pipeline
- VerTQ includes 5129 FP16/32 compute units
- Mapped to FPGA at 125 MHz, consumes 5.7 mm² in TSMC 16FF
- System produced four autonomous designs total
Entities
Institutions
- arXiv