Compute Units

Cube

Performs matrix computations.

• Each time the Cube is executed, a fp16 16 x 16 and 16 x 16 matrix multiplication is completed. For example, C = A x B. If int8 input is used, 16 x 32 and 32 x 16 matrix multiplication is completed at a time. A is from L0A, B is from L0B, and L0C stores the matrix multiplication result and intermediate result.
• Accumulator: adds the current matrix multiplication result to the result of the previous multiplication, implementing the addition of bias in convolution.

  For example, C = A x B + C.

• A/B DFF: data register for buffering the left or right submatrix multiplication result of the current 16 x 16 matrix.
• Accum DFF: data register for buffering the result 16 x 16 matrix.

The Cube Unit of different versions of Ascend AI Processors supported different data types, including fp16, fp32, uint8, int8, and int32.

Vector

Performs vector computations.

Compared with the Cube, the Vector offers less robust computing power but more flexible computations (such as the reciprocal and square root in mathematics). All source data and target data of Vector computations must be stored in the Unified Buffer and be 32-byte aligned.

The Vector Unit of different versions of Ascend AI Processors supported different data types, including fp16, fp32, int32, and int8.

Scalar

Computes scalar data of various types and controls the program flow.

It can be regarded as a small CPU, which implements loop control, branch judgment, address and parameter computations of Cube/Vector instructions, and basic arithmetic operations for a program.