High bandwidth memory system with crossbar switch for dynamically programmable distribution scheme

A method comprising:

• receiving a memory request provided from a first processing element at a first decomposition unit of a first memory unit, wherein the first memory unit includes a first plurality of memory banks and a first crossbar switch, the first crossbar switch communicatively connecting each of a plurality of decomposition units included in the first memory unit to each of the first plurality of memory banks included in the first memory unit, and wherein the first processing element is connected to a second memory unit including a second crossbar switch;
• decomposing the memory request for memory data spanning a plurality of data access units into at least a first partial request for a first subset of the memory data and a second partial request for a second subset of the memory data;
• determining that a requested first data of the first partial request resides in a first memory bank of the first plurality of memory banks;
• determining that a requested second data of the second partial request resides in a second memory bank of the first plurality of memory banks;
• directing the first partial request to the first memory bank via the first crossbar switch;
• directing the second partial request to the second memory bank via the first crossbar switch;
• retrieving the requested first data from the first memory bank via the first crossbar switch;
• retrieving the requested second data from the second memory bank via the first crossbar switch;
• preparing a partial response that includes the requested first and second data;
• and
• providing the partial response to the first processing element, wherein the first processing element is configured to access a plurality of memory units including the first memory unit and the second memory unit using a first dynamically programmable distribution scheme specifying a first programmatically modifiable distribution pattern of data of the first processing element for scattering data across the plurality of memory units, and wherein a second processing element is configured to access the plurality of memory units using a second dynamically programmable distribution scheme specifying a second programmatically modifiable distribution pattern of data of the second processing element for scattering data across the plurality of memory units.

A method comprising:

• receiving a first memory request provided from a first processing element at a first memory data request decomposition unit of a first memory unit, wherein the first memory unit includes a first plurality of memory banks and a first crossbar switch, the first crossbar switch communicatively connecting each of a plurality of decomposition units included in the first memory unit to each of the first plurality of memory banks included in the first memory unit, and wherein the first processing element is connected to a second memory unit including a second crossbar switch;
• receiving a second memory request provided from a second processing element at a second memory data request decomposition unit of the first memory unit;
• decomposing the first memory request for memory data spanning a first plurality of data access units into a first plurality of partial requests for a first subset of the memory data and the second memory request spanning a second plurality of data access units into a second plurality of partial requests for a second subset of the memory data;
• determining for each partial request of the first plurality of partial requests and the second plurality of partial requests whether the partial request is to be served from the first plurality of memory banks;
• discarding a first group of partial requests from the first plurality of partial requests and the second plurality of partial requests that is not to be served from the first plurality of memory banks;
• for each partial request of a second group of partial requests from the first plurality of partial requests that is to be served from the first plurality of memory banks, retrieving data of the partial request via the first crossbar switch, preparing a first partial response using the retrieved data, and providing the first partial response to the first processing element, wherein the first processing element is configured to access a plurality of memory units including the first memory unit and the second memory unit using a first dynamically programmable distribution scheme specifying a first programmatically modifiable distribution pattern of data for scattering data across the plurality of memory units;
• and
• for each partial request of a third group of partial requests from the second plurality of partial requests that is to be served from the first plurality of memory banks, retrieving data of the partial request via the first crossbar switch, preparing a second partial response using the retrieved data, and providing the second partial response to the second processing element, wherein the second processing element is configured to access the plurality of memory units using a second dynamically programmable distribution scheme specifying a second programmatically modifiable distribution pattern of data for scattering data across the plurality of memory units.