Final answer:
Two true statements about Nvidia CUDA are that not all threads in a grid execute in parallel due to hardware limitations, and a 2-dimensional grid can contain 3-dimensional blocks, providing flexibility in task organization. Other options are incorrect as thread blocks are included in grids and multiple blocks and warps can be assigned to an SM simultaneously.
Step-by-step explanation:
The question is about the Nvidia CUDA computing platform which allows developers to use a CUDA-enabled graphics processing unit (GPU) for general purpose processing (an approach known as GPGPU, General-Purpose computing on Graphics Processing Units). Particularly, the student is asking about the organization of threads, blocks, and warps within this architecture. There are two correct statements in the question:
- e) Not all threads in a grid actually execute in parallel. This is true because the number of threads that can execute in parallel is limited by the hardware resources such as the number of Stream Multiprocessors (SMs) and the hardware cap on the number of active threads per SM.
- c) A 2 dimensional grid can contain 3 dimensional blocks. This is true. In CUDA, grids can be organized into one, two, or three dimensions. Similarly, blocks within these grids can also be one, two, or three dimensional, providing a great deal of flexibility in how work is divided and organized.
The other options given to the student are incorrect due to the following reasons:
- a) Thread blocks are composed of threads and are part of a grid. The grid contains the blocks, not the other way around.
- b) Multiple blocks can be assigned to each SM on the device, depending on the resources required by those blocks and the available resources on the SM.
- d) Each SM can handle multiple warps concurrently. A warp is typically composed of 32 threads, and an SM will manage multiple warps at the same time.