Question

In the context of GPU occupancy, what happens to the number of active warps on a Streaming Multiprocessor when the register usage per thread is increased beyond the hardware limit?

Accepted Answer

When the register usage per thread increases beyond the hardware limit of a Streaming Multiprocessor, the kernel fails to launch entirely. Registers are high-speed memory locations inside the GPU used to store local variables for each thread. Each Streaming Multiprocessor has a fixed, finite pool of these registers that must be divided among all active threads. A Streaming Multiprocessor is the core processing unit of a GPU that executes warps, which are groups of thirty-two threads that execute the same instruction simultaneously. Because the hardware has a maximum number of registers available per thread, the compiler must allocate a specific amount of register space to every thread before execution begins. If your code requires more registers per thread than the hardware can physically provide, the GPU cannot accommodate the thread&#x27;s memory requirements, and the kernel execution is rejected by the driver. This is distinct from situations where register usage is merely high but within limits, which would simply reduce occupancy by limiting how many warps can fit on the chip at once. Exceeding the register limit is a hard boundary that prevents the workload from running.

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

In the context of GPU occupancy, what happens to the number of active warps on a Streaming Multiprocessor when the register usage per thread is increased beyond the hardware limit?

Community Answers