In this slide we see how memcpy can be performed entirely in memory. Rather than transfer each byte from memory to the processor and back, every DRAM line can be read into the row buffer and written back twice, once to the original row and again to the target row. This results in an effective copy rate thousands of times the original.
Given current memory architecture this would require memory to be copied in 16 Kb chunks aligned to 16 Kb since we can't just copy part of a row. In order to support smaller or unaligned copies, the memory could be built to have some sort of mask which allows partial rows to be written back.
This comment was marked helpful 0 times.
drayson
How much is this technique used in current hardware?
In this slide we see how memcpy can be performed entirely in memory. Rather than transfer each byte from memory to the processor and back, every DRAM line can be read into the row buffer and written back twice, once to the original row and again to the target row. This results in an effective copy rate thousands of times the original.
Given current memory architecture this would require memory to be copied in 16 Kb chunks aligned to 16 Kb since we can't just copy part of a row. In order to support smaller or unaligned copies, the memory could be built to have some sort of mask which allows partial rows to be written back.
This comment was marked helpful 0 times.
How much is this technique used in current hardware?
This comment was marked helpful 0 times.