SANDISK TECHNOLOGIES LLC

A computing system is provided that includes at least one processing unit, at least one high bandwidth memory (HBM) unit, and at least one high bandwidth flash (HBF) unit. The HBM and HBF units are all in electrical communication with the at least one processing unit. The computing system also includes control circuitry that is configured to train a large language model according to a low-rank adaptation (LoRA) technique. The control circuitry is configured to store a full-weight matrix in the at least one HBF unit and to store at least one low-rank matrix in the at least one HBM unit.

Anon-volatile memory structure capable of storing weights for layers of a deep neural network (DNN) and perform an inferencing operation within the structure is presented. An in-array multiplication can be performed between multi-bit valued inputs, or activations, for a layer of the DNN and multi-bit valued weights of the layer. Each bit of a weight value is stored in a binary valued memory cell of the memory array and each bit of the input is applied as a binary input to a word line of the array for the multiplication of the input with the weight. To perform a multiply and accumulate operation, the results of the multiplications are accumulated by adders connected to sense amplifiers along the bit lines of the array. The adders can be configured to multiple levels of precision, so that the same structure can accommodate weights and activations of 8-bit, 4-bit, and 2-bit precision.

Non-volatile memory structures for performing compute in memory inferencing for neural networks are presented. To improve performance, both in terms of speed and energy consumption, weight matrices are replaced with their singular value decomposition (SVD) and use of a low rank approximations (LRAs). The decomposition matrices can be stored in a single array, with the resultant LRA matrices requiring fewer weight values to be stored. The reduced sizes of the LRA matrices allow for inferencing to be performed more quickly and with less power. In a high performance and energy efficiency mode, a reduced rank for the SVD matrices stored on a memory die is determined and used to increase performance and reduce power needed for an inferencing operation.

Techniques are presented for performing in-memory matrix multiplication operations for binary input, binary weight valued convolution neural network (CNN) inferencing. The weights of a filter are stored in pairs of memory cells of a storage class memory device, such as a ReRAM or phase change memory based devices. To reduce current consumption, the binary valued filters are transformed into ternary valued filters by taking sums and differences of binary valued filter pairs. The zero valued weights of the transformed filters are stored as a pair of high resistance state memory cells, reducing current consumption during convolution. The results of the in-memory multiplications are pair-wise combined to compensate for the filter transformations. To compensate for zero valued weights, a zero weight register stores the number of zero weights along each bit line and is used to initialize counter values for accumulating the multiplication operations.

Use of a NAND array architecture to realize a binary neural network (BNN) allows for matrix multiplication and accumulation to be performed within the memory array. A unit synapse for storing a weight of a BNN is stored in a pair of series connected memory cells. A binary input is applied on a pair of word lines connected to the unit synapse to perform the multiplication of the input with the weight. The results of such multiplications are determined by a sense amplifier, with the results accumulated by a counter. The arrangement extends to ternary inputs to realize a ternary-binary network (TBN) by adding a circuit to detect 0 input values and adjust the accumulated count accordingly. The arrangement further extends to a ternary-ternary network (TTN) by allowing 0 weight values in a unit synapse, maintaining the number of 0 weights in a register, and adjusting the count accordingly.

Techniques are presented for accelerating in-memory matrix multiplication operations for a convolution neural network (CNN) inference in which the weights of a filter are stored in the memory of a storage class memory device, such as a ReRAM or phase change memory based device. To improve performance for inference operations when filters exhibit sparsity, a zero column index and a zero row index are introduced to account for columns and rows having all zero weight values. These indices can be saved in a register on the memory device and when performing a column/row oriented matrix multiplication, if the zero row/column index indicates that the column/row contains all zero weights, the access of the corresponding bit/word line is skipped as the result will be zero regardless of the input.

To improve efficiencies for inferencing operations of neural networks, ensemble neural networks are used for compute-in-memory inferencing. In an ensemble neural network, the layers of a neural network are replaced by an ensemble of multiple smaller neural network generated from subsets of the same training data as would be used for the layers of the full neural network. Although the individual smaller network layers are “weak classifiers” that will be less accurate than the full neural network, by combining their outputs, such as in majority voting or averaging, the ensembles can have accuracies approaching that of the full neural network. Ensemble neural networks for compute-in-memory operations can have their efficiency further improved by implementations based binary memory cells, such as by binary neural networks using binary valued MRAM memory cells. The size of an ensemble can be increased or decreased to optimize the system according to error requirements.

Non-volatile memory structures for performing compute-in-memory inferencing for neural networks are presented. A memory array is formed according to a crosspoint architecture with a memory cell at each crosspoint junction. The multi-levels memory cells (MLCs) are formed of multiple of ultra-thin dielectric layers separated by metallic layers, where programming of the memory cell is done by selectively breaking down one or more of the dielectric layers by selecting the write voltage level. In an alternate set of embodiments, the memory cells are formed as anti-fuses.

Non-volatile memory structures for performing compute-in-memory inferencing for neural networks are presented. A memory array is formed according to a crosspoint architecture with a memory cell at each crosspoint junction. The multi-levels memory cells (MLCs) are formed of multiple of ultra-thin dielectric layers separated by metallic layers, where programming of the memory cell is done by selectively breaking down one or more of the dielectric layers by selecting the write voltage level. In an alternate set of embodiments, the memory cells are formed as anti-fuses.

Non-volatile memory structures for performing compute in memory inferencing for neural networks are presented. To improve performance, both in terms of speed and energy consumption, weight matrices are replaced with their singular value decomposition (SVD) and use of a low rank approximations (LRAs). The decomposition matrices can be stored in a single array, with the resultant LRA matrices requiring fewer weight values to be stored. The reduced sizes of the LRA matrices allow for inferencing to be performed more quickly and with less power. In a high performance and energy efficiency mode, a reduced rank for the SVD matrices stored on a memory die is determined and used to increase performance and reduce power needed for an inferencing operation.