TensorRT
TensorRT copied to clipboard
NVIDIA
How will TRT utilize the optical flow accelerator?
the new announcement says that the Ada architecture will come with 300 OFA units that is dedicated to accelerate optical flow applications (according to the documentation, stereo matching can use it too). My question is, will TRT use the OFA units when we deploy OF or SM models? Or do we have to use the models specifically provided by NV’s SDK to utilize this amount of resources? Thanks!
I don't think so. @nvpohanh correct me if I'm wrong
excuse me, my question has two parts, which part do you not think so?
will TRT use the OFA units when we deploy OF or SM models? -> No, TRT doesn't support OFA. Or do we have to use the models specifically provided by NV’s SDK to utilize this amount of resources? -> Yes, but I don't know what the SDK is.
will TRT use the OFA units when we deploy OF or SM models? -> No, TRT doesn't support OFA. Or do we have to use the models specifically provided by NV’s SDK to utilize this amount of resources? -> Yes, but I don't know what the SDK is.
Thank you for the info! Do you have future plan to support OFA? It does seems like a big waste if not
I don't think so, TensorRT is mainly focused on GPU.
Ok, so that OFA unit is not like Tensor Core or something similar that is dedicated to certain type of calculation, but more like a "fpga" unit that is preprogramed to execute only certain models?
I am not able to answer this question since I know nothing about OFA :)
Ok I will just wait for the official release and see if there will be more documentation about OFA. Thanks!
We currently don't have plans to use OFA in TRT since it is mainly for gaming purpose, not DL.
Closing since no activity for more than 3 weeks, please reopen if you still have question, thanks!
