CUDA Questions and Answers for Viva

Answer-1: Multiple GPUs can be managed using device selection functions (cudaSetDevice) and separate streams for concurrent execution.

Answer-2: The CUDA driver API provides low-level access to GPU resources and enables more control over kernel execution and memory management than the runtime API.

Answer-3: Streams in CUDA allow concurrent execution of tasks like memory transfers and kernel execution, improving performance by overlapping operations.

Answer-4: PTX (Parallel Thread Execution) is an intermediate assembly-like language used by NVIDIA GPUs to represent parallel code.

Answer-5: cudaMemcpy is synchronous, blocking until the transfer is complete, while cudaMemcpyAsync is asynchronous, allowing the program to continue execution immediately.

Answer-6: The CUDA profiler is a tool for analyzing and optimizing the performance of CUDA applications, providing insights into memory usage, kernel execution, and bottlenecks.

Answer-7: CUDA (Compute Unified Device Architecture) is a parallel computing platform and programming model developed by NVIDIA for general-purpose computing on GPUs (GPGPU).

Answer-8: CUDA was developed by NVIDIA.

Answer-9: CUDA supports languages like C, C++, Fortran, and Python (via libraries like PyCUDA).

Answer-10: A CUDA kernel is a function written in CUDA C/C++ that runs on the GPU in parallel across multiple threads.

Answer-11: A thread is the smallest unit of execution in CUDA, responsible for executing a kernel function on a single data element.

Answer-12: A thread block is a collection of threads that execute the same kernel and can share memory and synchronize with each other.

Answer-13: A grid is a collection of thread blocks, enabling large-scale parallelism by organizing many threads across the GPU.

Answer-14: The GPU acts as the compute device, executing parallel tasks offloaded by the CPU.

Answer-15: Shared memory is a fast, on-chip memory shared among threads in the same block.

Answer-16: Global memory is the main memory on the GPU, accessible by all threads but slower than shared memory.

Answer-17: Local memory is memory private to a thread, stored in global memory, and used when registers are insufficient.

Answer-18: Constant memory is a small, read-only memory area accessible by all threads, optimized for broadcasting the same data across threads.

Answer-19: A CUDA kernel is defined using the __global__ keyword, e.g., __global__ void kernelName(...) { }.

Answer-20: CUDA kernels are launched with the syntax kernelName<<>>(arguments), where gridDim and blockDim specify the grid and block dimensions.

Answer-21: A warp is a group of 32 threads in CUDA that execute instructions in lockstep.

Answer-22: Coalesced memory access occurs when threads in a warp access consecutive memory addresses, improving memory access efficiency.

Answer-23: CUDA streams are sequences of operations (kernels, memory copies, etc.) that execute on the GPU in order or concurrently.

Answer-24: Synchronous operations block until completion, while asynchronous operations allow the host and device to execute other tasks concurrently.

Answer-25: GPU performance is typically measured using metrics like kernel execution time, memory throughput, and floating-point operations per second (FLOPS).

Answer-26: cudaMemcpy transfers data between the host and device memory.

Answer-27: cudaMemcpyHostToDevice transfers data from the host (CPU) to the device (GPU), while cudaMemcpyDeviceToHost transfers data in the opposite direction.

Answer-28: Pinned memory (page-locked memory) is a region of host memory that cannot be swapped out, enabling faster data transfers between host and device.

Answer-29: The __shared__ keyword declares shared memory that can be accessed by all threads within a block.

Answer-30: The __device__ keyword defines functions that are executed only on the GPU and callable only from other GPU functions or kernels.

Answer-31: __syncthreads() synchronizes all threads in a block, ensuring they reach the same point in execution before continuing.

Answer-32: Bank conflict occurs when multiple threads access the same memory bank, causing serialization of memory accesses and reducing performance.

Answer-33: A CUDA context is the environment within which CUDA kernels and memory operations execute, associated with a specific GPU.

Answer-34: CUDA provides error codes and helper functions like cudaGetErrorString() to report and interpret errors.

Answer-35: cudaMalloc allocates memory on the GPU device.

Answer-36: cudaFree deallocates memory previously allocated on the GPU.

Answer-37: CUDA device properties include characteristics like number of CUDA cores, shared memory size, global memory size, etc., retrievable using cudaGetDeviceProperties().

Answer-38: CUDA achieves parallelism by executing multiple threads in parallel on GPU cores, organized into blocks and grids.

Answer-39: __global__ functions are called from the host and run on the device, while __device__ functions are called only from the device and execute on the device.

Answer-40: Texture memory is a cached, read-only memory optimized for 2D spatial locality, often used in graphics or imaging applications.

Answer-41: Unified memory is a shared memory space accessible by both the CPU and GPU, simplifying memory management.

Answer-42: CUDA kernels can be optimized by improving memory access patterns, minimizing divergence, maximizing occupancy, and reducing shared memory bank conflicts.

Answer-43: Warp divergence occurs when threads in a warp follow different execution paths due to conditional statements, reducing performance.

Answer-44: The cudaEvent API is used to measure execution time, synchronize streams, and handle events between the host and device.

Answer-45: CUDA Thrust is a high-level C++ template library for parallel programming, simplifying operations like sorting, searching, and reductions on GPUs.

Answer-46: Dynamic parallelism allows GPU kernels to launch other kernels directly, enabling recursive algorithms and dynamic workloads.

Answer-47: The maximum number of threads per block is hardware-dependent, typically 1024 for modern GPUs.

Answer-48: Compute capabilities describe the features supported by a GPU, including maximum threads, memory sizes, and supported instructions.

Answer-49: Occupancy refers to the ratio of active warps to the maximum number of warps that can be run on an SM, indicating how well GPU resources are utilized.

Answer-50: CUDA atomic operations are operations that guarantee exclusive access to a memory location, preventing race conditions when multiple threads access the same memory.