Question

Explain the following CUDA terms: Global Memory, Shared Memory, Register, Host, Device, SM, SP, Grid, Block, and Thread.

Answer 1

Global memory:

• It is the computer storage which can be used by the number of processors which are connected together in order to make work as a multiprocessor system.
• In CUDA, global memory is the root for data access and each one of the accessing techniques starts at global memory.

Shared Memory:

• Shared memory is the memory which is being shared by the threads which can be in a thread block and will provide a mechanism for the threads to co-operate.
• The shared memory is faster the global memory as the access provided to it is faster than global memory.
• It is one of the most powerful features in CUDA for writing optimized codes.

Register:

• A register is known as a temporary storage area which is built in the CPU. The registers are the ones which can be used internally and also cannot be accessed outside the processors.
• There are many types of registers in CUDA. Some of them are as follows:
  • Instruction Register
  • Memory Data Register
  • Memory Buffer Register
  • Program Counter
  • Memory Address Register

Host:

• A host in CUDA is a device which communicated with the other hosts of the networks. There are a number of host in the network which communicates with each other.
• The TCP/IP network consists of the host which togetherly form a network identity and also forms the group of communications inside the network.

Device:

• The device is the term used for a particular technology which is made for adapting a particular purpose which can be of mechanical or electronic equipment.
• Hence, these are the device in CUDA.

SM & SP:

• Streaming Multiprocessors are the ones which are responsible for splitting into warps and here is the actual execution of the thread is performed with a greater cause.
• Stream Processors are used for allowing the application to more easily exploit the limited form of parallel processing.

Grid:

• A kernel launched as a collection of thread blocks is called a grid.
• It is a group of blocks. There is no synchronization between any of the blocks.

Block:

• A block is a group of threads. There is not much one can predict about the execution of the threads within a block.
• The block can execute concurrently or serially or is not any of the particular order.

Threads:

• They are the single execution units that are being run in the kernels on the GPU.
• The thread blocks are the collection of the threads.

Hence, these are all of the CUDA terms.