Back to results
US 2026/0180596 A1Published application
Adaptive quantization and compression for variable-length collective communication
AssigneesAdvanced Micro Devices, Inc., Xilinx, Inc.
InventorsVenkata Pavan Kumar Miriyala, Lucian Petrica, Kenneth O'Brien
Life of this patent
Estimating the term…
Abstract
A processing system dynamically and selectively quantizes or compresses variable-length input data for a collective operation to fit within a predetermined limit, executes the collective operation on the compressed data, and converts the data back to variable length results by dequantizing or decompressing the results of the collective operation.
Claims20
A method comprising:
- allocating a buffer having a fixed size for each process of a plurality of processes;
- selectively reducing a size of variable-length input data scheduled to be transferred between a first process of the plurality of processes and the other processes of the plurality of processes for a collective operation based on the variable-length input data exceeding the fixed size of the buffer;
- and
- transmitting the data to the buffer.
1 more element
A method, comprising:
- selectively reducing a precision of an aggregate of variable-length input data scheduled to be received by each process of a plurality of processes from other processes of the plurality of processes for a collective operation based on the aggregate of variable-length input data exceeding a predefined size limit;
- and
- transmitting the selectively reduced-precision input data to each process.
A processing system, comprising:
- a plurality of processes, each process of the plurality of processes configured to send variable-length input data to each of the other processes of the plurality of processes;
- and
- a network interface linking the processes configured to selectively reduce a precision of an aggregate of variable-length input data scheduled to be received by each process of the plurality of processes from the other processes of the plurality of processes for a collective operation based on the aggregate of variable-length input data exceeding a predefined size limit.
Dependencies are inferred from the claim text.