A technique that performs attention calculations (a core part of how transformers process information) using 8-bit floating-point numbers instead of higher precision, reducing memory and computation needs.