I’m working on an xavier agx / 32gb. The nvpmodel is set to MAXN (running jetson_clocks does not change the outcome).
returns 2133000000 (Hz) which is expected, and which together with the 256 bit bus width translates to the advertised 137GB/s memory bandwidth.
However, if I print the memoryClockRate obtained through cudaGetDeviceProperties the result is 1377000 (KHz) independently of the nvpmodel and/or jetson_clocks. So the implied memory bandwidth is not 137GB/s but only 88GB/s.
Given that CPU and GPU use the same physical memory, I would assume that the bus width and frequencies are the same.
Is there something wrong with the way I set the memory frequency + what is reported by /sys/kernel/debug/bpmp/debug/clk/emc/rate or is cudaGetDeviceProperties giving wrong results?
best regards, Lukas Wirz
Hi Lukas ,
Have you explored this page already on how to set frequency ?
2133000000 (Hz) looks correct for xavier agx 32gb.
1377000(Khz) looks like the GPU clock rate that you should get. What is the out of ./deviceQuery on you device ?
Yes, I know that page. It does not matter how I set the frequency.
deviceQuery unsurprisingly reports the same result I was seeing before:
root@x221-xavier:/usr/local/cuda/samples/1_Utilities/deviceQuery# echo 1 > /sys/kernel/debug/bpmp/debug/clk/emc/mrq_rate_locked
root@x221-xavier:/usr/local/cuda/samples/1_Utilities/deviceQuery# echo 1 > /sys/kernel/debug/bpmp/debug/clk/emc/state
root@x221-xavier:/usr/local/cuda/samples/1_Utilities/deviceQuery# echo 1866000000 > /sys/kernel/debug/bpmp/debug/clk/emc/rate
root@x221-xavier:/usr/local/cuda/samples/1_Utilities/deviceQuery# cat /sys/kernel/debug/bpmp/debug/clk/emc/rate
CUDA Device Query (Runtime API) version (CUDART static linking)
Detected 1 CUDA Capable device(s)
Device 0: "Xavier"
CUDA Driver Version / Runtime Version 10.2 / 10.2
CUDA Capability Major/Minor version number: 7.2
Total amount of global memory: 31929 MBytes (33479491584 bytes)
( 8) Multiprocessors, ( 64) CUDA Cores/MP: 512 CUDA Cores
GPU Max Clock rate: 1377 MHz (1.38 GHz)
Memory Clock rate: 1377 Mhz
Memory Bus Width: 256-bit
I suspect , cudaDeviceProp::memoryClockRate might be broken on Jetson , Let me investigate.
For memory bandwidth you should refer to /sys/kernel/debug/bpmp/debug/clk/emc/rate
Are there any updates about that?
cudaDeviceProp::memoryClockRate is reporting wrong value on Jetson . We are targeting a fix in the upcoming release
Please rely on /sys/kernel/debug/bpmp/debug/clk/emc/rate for memory bandwidth
This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.