Neural network accelerator

1. A method for implementing a neural network, the method comprising:

receiving input data;

fetching, from a memory, weights of the neural network;

performing a first portion of processing for the neural network, the first portion implemented in hardware by an accelerator including a plurality of parallel multiply and accumulate (MAC) units configured to perform a plurality of MAC operations to generate a first neuron value at a first accumulator and a second neuron value at a second accumulator, wherein the first portion includes:

receiving a first subset of the input data from a circular buffer at inputs of the plurality of parallel MAC units;

performing the plurality of parallel MAC operations using a first set of weights, while holding the inputs of the plurality of parallel MAC units stable at the first subset of the input data, to generate a first portion of the first neuron value at the first accumulator;

performing the plurality of parallel MAC operations using a second set of weights, while holding the inputs of the plurality of parallel MAC units stable at the first subset of the input data, to generate a second portion of the second neuron value at the second accumulator;

repeating receiving subsets of input data from the circular buffer and performing the plurality of parallel MAC operations while holding the inputs of the plurality of parallel MAC units stable to accumulate the first neuron value at the first accumulator and the second neuron value at the second accumulator for all input data;

selecting the first neuron value before a bias and an activation function is applied, using a multiplexer coupled to the plurality of parallel MAC units; and

writing the first neuron value to the memory; and

performing a second portion of processing for the neural network, the second portion implemented in software by a processor, the accelerator and the processor using a bus to communicate and to share access to the memory, wherein the second portion includes:

reading the first neuron value from the memory; and

applying the bias and the activation function to the first neuron value.

2. The method according to claim 1 , wherein the first set of weights and the second set of weights are compressed weights.

3. The method according to claim 2 , wherein the first portion of processing includes decompressing the compressed weights using a lookup table to generate decompressed weights.

4. The method according to claim 3 ,

wherein:

the compressed weights are stored in the memory and are each 4 bits; and

the decompressed weights are used in the first portion of the processing of the neural network and are each 8 bits.

5. The method according to claim 4 , wherein the compressed weights stored in the memory are pruned so that zero value weights are encoded as a single zero bit and non-zero value weights are encoded with a header bit of one.

6. The method according to claim 3 , wherein the lookup table is implemented in software running on the processor and loaded into the accelerator.

7. The method according to claim 6 , wherein the lookup table is implemented in hardware on the accelerator.

8. The method according to claim 2 , wherein the first set of weights and the second set of weights include pruned weights and unpruned weights, the pruned weights including single bit representations of synapses and the unpruned weights including an additional bit.

9. The method according to claim 1 , wherein the plurality of parallel multiply and accumulate operations are performed by a plurality of MAC units operating in parallel and repetitively to produce a sum of products corresponding to each neuron in the neural network.

10. The method according to claim 1 , wherein the memory and the accelerator are part of a hearing aid.