Question

What specific memory access pattern must be satisfied by a warp to ensure that all global memory requests are serviced in a single transaction by the memory controller?

Accepted Answer

To ensure that all global memory requests from a warp are serviced in a single transaction, the memory access pattern must be coalesced. A coalesced access occurs when all 32 threads in a warp access a contiguous, aligned block of memory. The memory address accessed by the warp must be aligned to the size of the memory transaction, which is typically 128 bytes. This means that the base address accessed by the warp divided by the size of the transaction must result in an integer. Furthermore, all threads must access data of the same size, such as 4-byte integers or 8-byte doubles, within that single aligned 128-byte segment. If the threads access memory addresses that fall within this single 128-byte window, the hardware memory controller can combine these individual requests into one efficient memory transaction, minimizing the time the processor waits for data. If threads access addresses outside of this aligned window or access non-contiguous memory, the hardware must issue multiple separate memory transactions to fetch the required data, which significantly reduces overall performance.

Home → All Courses → Engineering and Technology Courses → High-Performance GPU Programming for AI Workloads → Flashcard

What specific memory access pattern must be satisfied by a warp to ensure that all global memory requests are serviced in a single transaction by the memory controller?