SIMD < SIMT < SMT: Parallelism in Nvidia GPUs (2011)

The content discusses NVIDIA GPUs and their programming models, focusing on SIMT, SIMD, and SMT. The SIMT model is a hybrid between vector processing and hardware threading, offering both flexibility and efficiency. SIMT is a more flexible version of SIMD, with SMT being even more flexible, but less efficient. The content highlights the benefits and costs of SIMT, including parallel random access and handling flow divergence. Comparisons are made between SIMT and SMT, where SMT is seen as an afterthought for processors originally designed for single-threaded performance. The article delves into the intricacies of GPU hardware design and the trade-offs involved in different parallel programming models.

https://yosefk.com/blog/simd-simt-smt-parallelism-in-nvidia-gpus.html

To top