Multiply Accumulate Unit Part-I

MultiplyAccumulateUnit

A stateful arithmetic module that repeatedly forms A×B and adds the selected product into persistent Sum state.

MAC

Abbreviation for multiply-accumulate: an operation that combines multiplication with repeated accumulation.

Sum_next = Sum_current + (A × B)

The fundamental state-update equation of a multiply-accumulate unit.

MAC Product Formation

The combinational stage that interprets A and B, multiplies them, and produces the product that will later be accumulated.

parameter WIDTH = 8

Defines the bit width of input operands A and B in MultiplyAccumulateUnit.

parameter ACC_WIDTH = 2 * WIDTH

Defines the accumulator width, defaulting to the natural full-product width of two WIDTH-bit operands.

input wire [WIDTH - 1:0] A

The WIDTH-bit raw bit pattern of the first multiplication operand.

input wire [WIDTH - 1:0] B

The WIDTH-bit raw bit pattern of the second multiplication operand.

input wire SignedMode

Selects whether the MAC interprets A and B through the signed or unsigned multiplication path.

wire [(2 * WIDTH) - 1:0] Product

The selected 2×WIDTH-bit product that will subsequently be extended to the accumulator width.

wire signed [WIDTH - 1:0] SignedA

A signed Verilog view of the WIDTH-bit pattern carried by A.

wire signed [WIDTH - 1:0] SignedB

A signed Verilog view of the WIDTH-bit pattern carried by B.

wire signed [(2 * WIDTH) - 1:0] SignedProduct

The 2×WIDTH-bit result produced by multiplying SignedA and SignedB using signed interpretation.

wire [(2 * WIDTH) - 1:0] UnsignedProduct

The 2×WIDTH-bit result produced by multiplying A and B using unsigned interpretation.

SignedA and A contain the same bits.

Does assign SignedA = A numerically transform or rearrange the bits of A?

SignedB and B contain the same bits.

Does assign SignedB = B numerically transform or rearrange the bits of B?

assign SignedA = A;

Copies A's bit pattern into a signal explicitly declared signed.

assign SignedB = B;

Copies B's bit pattern into a signal explicitly declared signed.

The bit pattern is preserved while its arithmetic interpretation changes.

What is the important effect of assigning A to SignedA and B to SignedB?

Because Verilog multiplication behavior depends on the signedness of the operands, so explicitly signed signals provide the intended signed interpretation.

Why does MultiplyAccumulateUnit create SignedA and SignedB?

A and B

Operands used by the unsigned multiplication path.

SignedA and SignedB

Operands used by the signed multiplication path.

assign UnsignedProduct = A * B;

Computes the product using the unsigned interpretations of A and B.

assign SignedProduct = SignedA * SignedB;

Computes the product using the signed interpretations of the same operand bit patterns.

UnsignedProduct = A × B interpreted unsigned

What arithmetic interpretation produces UnsignedProduct?

SignedProduct = SignedA × SignedB interpreted signed

What arithmetic interpretation produces SignedProduct?

Both product paths exist concurrently in the RTL.

Does SignedMode determine whether the signed or unsigned multiplication expression physically exists in the source code?

Because the module separately assigns both UnsignedProduct and SignedProduct and selects between their results afterward.

Why can the RTL be viewed as forming both signed and unsigned product candidates?

assign Product = (SignedMode) ? SignedProduct : UnsignedProduct;

Selects which product interpretation continues through the MAC datapath.

SignedMode = 0

Product receives UnsignedProduct.

SignedMode = 1

Product receives SignedProduct.

Product Selection

The stage in which SignedMode chooses between SignedProduct and UnsignedProduct.

SignedMode acts after the two product candidates are formed.

Where does SignedMode directly select between the multiplication results in the shown RTL?

SignedMode ? SignedProduct : UnsignedProduct

The ternary expression implementing signed-versus-unsigned product selection.

Product is 2×WIDTH bits wide.

How wide is the selected Product before it is extended to ACC_WIDTH?

Because multiplying two WIDTH-bit operands can require up to 2×WIDTH bits to preserve the full product representation.

Why are SignedProduct, UnsignedProduct, and Product all declared 2×WIDTH bits wide?

Product Width and Accumulator Width Are Separate Questions

The architectural principle that the multiplication result has its own natural width while persistent Sum may use a different width.

2×WIDTH

Product width produced by the MAC's multiplication stage.

ACC_WIDTH

Width used by the MAC's persistent accumulator state.

ACC_WIDTH = 2×WIDTH by default.

What relationship exists between product width and accumulator width under the module's default parameters?

ACC_WIDTH may be wider than 2×WIDTH.

Can the MAC accumulator be configured wider than the multiplication result?

Because the product represents one A×B operation, whereas Sum may need to retain the accumulated result of many products.

Why might a MAC require an ACC_WIDTH greater than its 2×WIDTH product width?

Numeric Interpretation

The meaning assigned to a bit pattern when arithmetic hardware treats it as signed or unsigned.

The same bit pattern can represent different numeric values depending on signedness.

Why is signedness an architectural concern even when the physical operand bits do not change?

A, B → UnsignedProduct

The unsigned interpretation path through the multiplication stage.

A, B → SignedA, SignedB → SignedProduct

The signed interpretation path through the multiplication stage.

SignedProduct or UnsignedProduct → Product

The convergence point of the two numeric-interpretation paths.

Signed MAC Path

Interprets A and B as signed values, forms SignedProduct, and selects it when SignedMode is asserted.

Unsigned MAC Path

Interprets A and B as unsigned values, forms UnsignedProduct, and selects it when SignedMode is deasserted.

The signed MAC does not merely change the final output label.

Is SignedMode simply metadata describing Product after multiplication?

Because signedness determines how the operand bit patterns are interpreted during multiplication and later how the resulting Product must be widened for accumulation.

Why does SignedMode affect more than just the final Product label?

Signed interpretation must remain consistent through multiplication and extension.

What requirement follows after the MAC selects a signed Product?

Unsigned interpretation must remain consistent through multiplication and extension.

What requirement follows after the MAC selects an unsigned Product?

Signed product → sign extension

Which widening operation must eventually be applied when the selected Product represents a signed two's-complement value?

Unsigned product → zero extension

Which widening operation must eventually be applied when the selected Product represents an unsigned value?

Because widening must preserve the numeric meaning established by the selected multiplication interpretation.

Why must product extension agree with SignedMode?

SignedMode coordinates an arithmetic path, not merely one multiplier symbol.

What larger architectural lesson does the signed/unsigned MAC demonstrate?

Operand interpretation → multiplication → product selection

The first three conceptual stages of the MultiplyAccumulateUnit datapath.

Product selection is not yet accumulation.

Has the persistent Sum state been updated merely because Product has been selected?

Because Product is still a combinational multiplication result; it must next be extended appropriately and added into Sum on an enabled clock edge.

Why is the selected Product only an intermediate value in the MAC?

Multiplication is the transformation stage of the MAC.

Within the MAC's larger iterative pattern, what role does A×B perform before the result is combined with state?

Product

A combinational transformation of the current A and B operands selected according to SignedMode.

Sum

Persistent arithmetic history containing products accepted during previous enabled updates.

Product is current-data arithmetic; Sum is retained arithmetic history.

What is the key architectural distinction between Product and Sum in MultiplyAccumulateUnit?

A MAC is multiplication plus memory.

What distinguishes the MultiplyAccumulateUnit from a module that merely computes A×B?

Because the product is ultimately folded into a retained Sum that participates in later MAC operations.

Why is multiplication alone insufficient to define a MAC?