Update readme #306
Update readme #306
Conversation
tb_plugin/docs/gpu_utilization.md
Outdated
| For example, a kernel with only one thread per block can’t fully utilize each SM. | ||
|
|
||
| * Est. Achieved Occupancy: The bigger, the better. The definition of occupancy is [here](https://docs.nvidia.com/gameworks/content/developertools/desktop/analysis/report/cudaexperiments/kernellevel/achievedoccupancy.htm). | ||
| * Est. Achieved Occupancy: For most cases such as memory bandwidth bounded kernels, the higher the better. The definition of occupancy is [here](https://docs.nvidia.com/gameworks/content/developertools/desktop/analysis/report/cudaexperiments/kernellevel/achievedoccupancy.htm). |
There was a problem hiding this comment.
lets simply call it memory bounded instead of memory bandwidth bounded
There was a problem hiding this comment.
I am actually not quite sure how to make it clear in the case of occupancy. The core logic behind occupancy is to over-subscribe the GPU to hide the latency, since intra-warp context switch is quite fast. Once target occupancy is achieved, it does not make it better after that.
See page 13 of GPU Performance Analysis and Optimization
I think we should remove the whole sentence here.
There was a problem hiding this comment.
Memory bounded could be interpreted as size bounded or bandwidth bounded. So I prefer to add "bandwidth".
tb_plugin/docs/gpu_utilization.md
Outdated
| For example, a kernel with only one thread per block can’t fully utilize each SM. | ||
|
|
||
| * Est. Achieved Occupancy: The bigger, the better. The definition of occupancy is [here](https://docs.nvidia.com/gameworks/content/developertools/desktop/analysis/report/cudaexperiments/kernellevel/achievedoccupancy.htm). | ||
| * Est. Achieved Occupancy: For most cases such as memory bandwidth bounded kernels, the higher the better. The definition of occupancy is [here](https://docs.nvidia.com/gameworks/content/developertools/desktop/analysis/report/cudaexperiments/kernellevel/achievedoccupancy.htm). |
There was a problem hiding this comment.
I am actually not quite sure how to make it clear in the case of occupancy. The core logic behind occupancy is to over-subscribe the GPU to hide the latency, since intra-warp context switch is quite fast. Once target occupancy is achieved, it does not make it better after that.
See page 13 of GPU Performance Analysis and Optimization
I think we should remove the whole sentence here.
|
@cloudhan Cool reference URL. I've added it after this sentence as Reference. Giving user more info to judge is good. |
3 participants
No description provided.