I’m an engineering student who has just started to Cuda!
I would like to have some information:

-main differences between SIMT and SIMD

-while a kernel grid runs on 2 SM (for example) ,another kernel grid may be executed in parallel on other SM?

The CUDA programming guide has a short overview of SIMT in Chapter 4. Do you have a specific question about SIMT?

No. Currently, CUDA gives a kernel grid total control over all SMs until the kernel is finished executing. Two kernels cannot run at the same time.

the same kernel is assigned to warp?

therads of the wapr are performed independently in the SMs?

thanks for your attention

The main feature of SIMT is that the hardware automatically handles thread divergence with the warp.

All warps run code from the same kernel. There is one program counter per warp.

The 32 threads of the warp are executed on the 8 processors of the multiprocessor over 4 clock cycles.

BTW, this recent paper is very good for understanding the details of GPU hardware utilization:…09hpg_paper.pdf