BitNet Transformer Optimisation for FPGA

LLM BitNet Ternary Weights Lookup Tables FPGA

This project explores efficient inference for Microsoft’s BitNet b1.58 2B4T transformer. I converted its 2.08 billion ternary parameters into lookup-table operations designed for FPGA implementation.

What I Built

• Extracted and validated ternary weights across all 30 transformer layers
• Converted seven linear projections per layer into lookup-table representations
• Built a full inference emulator for checking the converted model
• Confirmed token-for-token agreement between lookup-table and direct ternary inference
• Eliminated DSP use for the main linear products
• Increased estimated throughput by 1.83× by grouping four ternary weights per lookup

Tech Stack

Python, PyTorch, BitNet b1.58, ternary neural networks, lookup tables, fixed-point arithmetic and FPGA-oriented inference.