Questions about tx2 GPU compute capability

Robotics & Edge Computing Jetson & Embedded Systems Jetson TX2
1

Hi Everyone:
We want to know the compute performance of tx2 GPU theoretically.But there are some pieces of information confused us.

1.According to https://en.wikipedia.org/wiki/Tegra#Devices_6.
	We know that tx2 gpu maximum working freqency is 1465 MHz and the FP16 compute capability is 1500 GFLOPS

2.According to Jetson_TX2_TX2i_Module_DataSheet_v01.pdf
	The GPU maximum operating frequency is 1.12GHZ.We assume the FP16 compute capability is: SMs*FP16 unit pre SM*2*clock rate = 2*256*2*1.12=1146.88 GFLOPS

3.cat /sys/devices/17000000.gp10b/devfreq/17000000.gp10b/available_frequencies
	We found the maximum clock freq could be 1300500000,but we can only set 1122000000 as the maxinum freq, the 1300500000 is a invalid argument.

Now we have the following questions:

Q1:Is the maximum frequency of TX2 GPU on Wiki wrong?
Q2:Is the arithmetic SMs*FP16 unit pre SM*2*clock rate = 2*256*2*1.12=1146.88 GFLOPS right?
Q3:Why we couldn't set 1300500000 as GPU frequency ?
2

The wikipedia.org article may be citing the TX2 chip as opposed to Jetson TX2, which can have a different maximum due to factors like cooling, TDP, MTBF, ect. See the Jetson community wiki on eLinux.org here: [url]https://elinux.org/Jetson[/url]

Try setting the max clocks with “sudo nvpmodel -m 0”. This will set the maximum GPU frequency to 1300500000 (I just double-checked on my Jetson TX2 with JetPack 3.3). The DVFS governor will still be enabled, so it won’t be always be running at max clocks unless you run a GPU-intensive application or if you run “~/jetson_clocks.sh”. The jetson_clocks script disables the governor and pegs the clocks at their maximum always (and also turns the fan on).

The correct peak GFLOPS is 22562*1.3005=1331.71 GFLOPS

3

You can also refer to “Max-Q and Max-P Power Efficiency” section in https://developer.nvidia.com/embedded/dlc/l4t-documentation-28-2-ga for max gpu frequency in different modes.