Hello,
I couldn’t reach the BIOS settings. I wonder if the SMI handlers enabled or not ?
Thanks!
Embedded systems don’t have a BIOS…software such as U-Boot are customized to take over that function. SMI functions most likely also depend on a PCI interface and the integrated GPU of all current Jetson embedded systems are directly connected to the memory controller without going through PCI. So none of that will be available from a TX2.
Thank you linuxdev,
I have made hwlatdetect test and there were some big latencies. So, I thought SMI cause it. Do you have any idea what is cause these latancies ?
By default the Jetsons are shipped set to not use high performance, but to instead use an energy efficiency mode. The first thing to do before any benchmark or performance test would be to put the Jetson in performance mode. Run:
sudo nvpmodel -m 0
sudo ~nvidia/jetson_clocks.sh
“nvpmodel” sets up some tables for clocks and voltages, “-m 0” allows the clocks and voltages to go to full performance. “jetson_clocks.sh” without any arguments will put the clocks to the max of the tables.
If performance still has issues after those settings, then you have good test data and can go from there.
I am doing all my test in mode 0 and with the clock. But Interesting part is that for this hwlatdetect test, mode 0 results are much more worse then mode 3’s.
Hwlatdetect is a program that controls the kernel hardware latency detector module (hwlat_detector.ko). The module is a special purpose kernel module that is used to detect large system latencies induced by the behavior of certain underlying hardware or firmware, independent of Linux itself.
I don’t understand why mode 0 is worse and unstable for these system latencies performance.
Do keep in mind that if you use a GPU there is a startup cost and the batch size will have a large impact. Initial latency will go up, but as the batch size goes up speed won’t suffer. I don’t know enough to give you much advise in that area, but perhaps having the Denver cores online has an additional latency effect. Once the pipeline starts you can get a lot of data throughput without latency suffering. You probably need to provide more details about your test, and then someone who knows more about it could comment.