A specialized 8-bit floating-point format designed for efficient neural network computation while maintaining reasonable accuracy.