Accelerating AI by Learning Optimal Datatypes​

At the heart of ByteShape’s acceleration stack is a suite of proprietary datatype learning algorithms that automatically determine the minimal precision required for each neural-network weight and activation. Unlike static quantization, ShapeLearn performs dynamic precision allocation to preserve accuracy while greatly reducing arithmetic complexity, memory footprint, and energy consumption.

ShapeLearn adapts to a wide range of deployment scenarios:

• Learns datatypes at any granularity — block, tensor, channel, or group
• Supports multiple precision types: integer, floating-point, and MX
• Integrates easily with existing quantization workflows
• Aligns learned precisions with native hardware numeric formats

When deployed on GPUs, FPGAs, or custom ASICs, models optimized with ShapeLearn achieve up to 10× faster inference, 7× faster training, and meaningful reductions in carbon footprint.

Complementing ShapeLearn, the ShapeSqueeze lossless compression engine applies per-value entropy coding on top of learned precisions, providing up to 40% additional reduction in off-chip memory traffic. The result: lower bandwidth pressure, reduced latency, and higher throughput.

Together, ShapeLearn and ShapeSqueeze enable large-scale LLMs, computer-vision pipelines, and edge deployments to run with exceptional efficiency.