Figure 7

Data transfer with the bus lock. A new (global) variable called lock is introduced. This variable (or a flag) is shared across all the GPUs sharing the common data bus so that each time a GPU needs to send or receive data, it waits until the bus is free, then it locks the bus, transfers data, and finally unlocks the bus. (a) Single GPU, specific sizes. (b) Two GPUs, specific sizes. (c) Single GPU, arbitrary sizes. (d) Two GPUs, arbitrary sizes.