How to set Target Nvidia A6000

Questions
#1

Is It enough to set tvm.target.Target(‘cuda’) in order for TVM to take full advantage of specific hardware platforms? because I have seen targets like “nvidia/geforce-rtx-3090”, “nvidia/nvidia-t4” and so on, but it is not a right target for “nvidia/geforce-rtx-a6000”. When I use target "nvidia/geforce-rtx-3090, the target specific information like this:

cuda -keys=cuda,gpu -arch=sm_86 -max_num_threads=1024 -max_threads_per_block=1024 -registers_per_block=65536 -shared_memory_per_block=49152 -thread_warp_size=3,

but if I set the target “cuda -arch=sm_86”, the target specific information only such like:

cuda -keys=cuda,gpu -arch=sm_86 -max_num_threads=1024 -thread_warp_size=32

should I add all other arget specific information for getting better tuning performance? Or could TVM get these information by it self?

#2

Yeah, tvm does not have A6000 target currently. I think you can add the A6000 target by customize tvm/tag.cc. But use rtx-3090 config for A6000 tuning will not decrease the performance in my point of view since they have the same compute core and the only difference is that A6000 has more onchip memory.

2 Likes
#3

Thanks for noticing the missing target tag. I agree with Lei that this should be the same as rtx-3090 config. For your other question on missing target attributes you may need to add other target attributes, it could impact later usage in the tuning system.

Also please feel free to send a PR to add this target tag.

#4

I created a PR to add this target tag, and added source for future reference. https://github.com/apache/tvm/pull/14762

2 Likes