Dynamic memory traffic optimization in multi-client systems

A method comprising:

• storing, in a buffer, data retrieved from a memory subsystem;
• receiving, by a client device, a memory access request from a requesting agent targeting a first data portion of a plurality of data portions;
• accessing, by the client device, metadata associated with the plurality of data portions;
• in response to determining from the metadata that the buffer stores a second data portion of the plurality of data portions that can be used to supply the first data portion, wherein the second data portion is different from the first data portion:
• preventing sending, by the client device, the memory access request to the memory subsystem;
• and
• generating, by the client device, the first data portion from the second data portion stored in the buffer;
• and
• sending, by the client device, the first data portion to the requesting agent.

A system comprising:

• a plurality of client devices;
• a memory subsystem;
• and
• a communication fabric coupled to the plurality of client devices and the memory subsystem;
• wherein a given client device of the plurality of client devices is configured to:
• receive a memory access request from a requesting agent targeting a first data portion of a plurality of data portions;
• access metadata associated with the plurality of data portions;
• in response to determining from the metadata that a local buffer stores a second data portion of the plurality of data portions that can be used to supply the first data portion, wherein the second data portion is different from the first data portion:
• prevent sending the memory access request to the memory subsystem;
• generate the first data portion from the second data portion stored in the local buffer;
• and
• send the first data portion to the requesting agent.