Neural network accelerator
Disclosed is a neural network accelerator including a first bit operator generating a first multiplication result by performing multiplication on first feature bits of input feature data and first weight bits of weight data, a second bit operator generating a second multiplication result by performing multiplication on second feature bits of the input feature data and second weight bits of the weight data, an adder generating an addition result by performing addition based on the first multiplication result and the second multiplication result, a shifter shifting a number of digits of the addition result depending on a shift value to generate a shifted addition result, and an accumulator generating output feature data based on the shifted addition result.
1. A neural network accelerator comprising:
a first bit operator configured to generate a first multiplication result by performing multiplication on first feature bits of input feature data and first weight bits of weight data;
a second bit operator configured to generate a second multiplication result by performing multiplication on second feature bits of the input feature data and second weight bits of the weight data;
an adder configured to generate an addition result by performing addition based on the first multiplication result and the second multiplication result;
a shifter configured to shift a number of digits of the addition result depending on a shift value to generate a shifted addition result; and
an accumulator configured to generate output feature data based on the shifted addition result,
wherein a number of digits of the first multiplication result and a number of digits of the second multiplication result are same.
2. The neural network accelerator of claim 1 , wherein the shift value is determined based on a number of digits of the first feature bits in the input feature data and a number of digits of the first weight bits in the weight data or is determined based on a number of digits of the second feature bits in the input feature data and a number of digits of the second weight bits in the weight data.
3. The neural network accelerator of claim 1 , wherein the input feature data is divided into the first feature bits and the second feature bits depending on a number of operating bits of each of the first bit operator and the second bit operator, and
wherein the weight data is divided into the first weight bits and the second weight bits depending on the number of operating bits.
4. The neural network accelerator of claim 1 , wherein a number of bits of the first feature bits is identical to a number of bits of the first weight bits.
5. The neural network accelerator of claim 1 , wherein the first bit operator generates the first multiplication result based on one selected depending on a selection signal among an XNOR operation and an AND operation.
6. The neural network accelerator of claim 1 , wherein the output feature data is a multiplication result for the input feature data and the weight data.
7. A neural network accelerator comprising:
a first processing circuit configured to generate a first operation result by adding a first multiplication result for first feature bits of input feature data and first weight bits of weight data and a third multiplication result for third feature bits of the input feature data and third weight bits of the weight data;
a second processing circuit configured to generate a second operation result by adding a second multiplication result for second feature bits of the input feature data and second weight bits of the weight data and a fourth multiplication result for fourth feature bits of the input feature data and fourth weight bits of the weight data;
a first shifter configured to shift a number of digits of the first operation result depending on a first shift value to generate a first shifted operation result;
a second shifter configured to shift a number of digits of the second operation result depending on a second shift value different from the first shift value to generate a second shifted operation result; and
an accumulator configured to generate output feature data based on the first shifted operation result and the second shifted operation result,
wherein a number of digits of the first multiplication result and a number of digits of the third multiplication result is same, and a number of digits of the second multiplication result and a number of the fourth multiplication result are same.
8. The neural network accelerator of claim 7 , wherein the first shift value is determined based on a number of digits of the first feature bits in the input feature data and a number of digits of the first weight bits in the weight data, and
wherein the second shift value is determined based on a number of digits of the second feature bits in the input feature data and a number of digits of the second weight bits in the weight data.
9. The neural network accelerator of claim 7 , wherein a number of bits of the first feature bits is identical to a number of bits of the first weight bits.
10. The neural network accelerator of claim 7 , wherein the output feature data is a multiplication result for the input feature data and the weight data.