Back to results
US 2026/0104966 A1Published application
Fast checkpointing of neural network training
AssigneeNVIDIA Corporation
InventorsSzymon Józef Karpinski, Kamil Tokarski
Life of this patent
Estimating the term…
Abstract
Apparatuses, systems, and techniques to perform fast checkpointing of training states during training of a neural network. In at least one embodiment, a checkpoint is created using a memory-to-memory copy operation that copies training state information from a graphics processing unit (GPU) memory to a pinned shared region of a central processing unit (CPU) memory, without creating an additional copy of this information in a process memory allocated to a training process.
Claims20
A processor comprising:
- one or more circuits to cause one or more pinned shared memory buffers to be allocated to one or more neural network accelerators, wherein the one or more pinned shared memory buffers are to store a plurality of neural network training checkpoints to be used by the one or more neural network accelerators to recover from one or more errors detected during training of one or more neural networks.
A method, comprising:
- causing one or more pinned shared memory buffers to be allocated to one or more neural network accelerators, wherein the one or more pinned shared memory buffers are to store a plurality of neural network training checkpoints to be used by the one or more neural network accelerators to recover from one or more errors detected during training of one or more neural networks.
A system, comprising:
- one or more computer devices to cause one or more pinned shared memory buffers to be allocated to one or more neural network accelerators, wherein the one or more pinned shared memory buffers are to store a plurality of neural network training checkpoints to be used by the one or more neural network accelerators to recover from one or more errors detected during training of one or more neural networks.
Dependencies are inferred from the claim text.