GB20C Virtualization - G-die Statistics

Purpose and Use Cases

NvGPU exposes certain control knobs for Linux to allow apps running on CPU side to acquire G-die statistics related to iGPU IPs, including GR3D, NVENC, NVDEC and OFA engines. This is similar to functionalities provided by the tegrastats tool. However, you can build DriveOS your own app by querying these G-die statistics. In particular, the rendered iGPU engine load info is generated at the engine level. Namely, even the statistics are queried by one VM, and the returned result is still combined utilization across all active VMs. Typical usage of this engine level utilization info can be for debug purpose or for hardware QA power performance estimation. NVIDIA does not recommend using this information to adjust engine clocks dynamically because iGPU IP’s clock setting is not expected to change in an automotive product.

Related Control Knobs for Linux VM

• iGPU Hardware Engine Utilization

  Name	Access Path	Functionality	Unit
  GR3D Engine Utilization	/sys/devices/platform/22100000.pcie/pci0000:01/0000:01:00.0/load	Returns an integer value as percentage of engine clock cycles where GR3D engine remains busy, across 2 GPCs.	%
  NVENC Engine Utilization	/sys/devices/platform/22100000.pcie/pci0000:01/0000:01:00.0/nvdec0_load	Returns an integer value as percentage of engine clock cycles where NVDEC engine remains busy.	%
  NVDEC Engine Utilization	/sys/devices/platform/22100000.pcie/pci0000:01/0000:01:00.0/nvenc0_load	Returns an integer value as percentage of engine clock cycles where NVENC engine remains busy.	%
  OFA Engine Utilization	/sys/devices/platform/22100000.pcie/pci0000:01/0000:01:00.0/ofa_load	Returns an integer value as percentage of engine clock cycles where OFA engine remains busy.	%

  • Suggested Query Method

    • 17ms query interval or similar value, based on 60 frames per second calculation.

    • Optional moving average over most recent X samples to smooth the curve, e.g., smooth over 10 samples:

      average_load = current_load * 0.9 + new_query * 0.1
      

    • Print (averaged) load value to user interface every 1s or larger interval.

  • Query interval must be greater than or equal to 10ms, considering ~3ms message round-trip time and to avoid overloading system.

• iGPU Engine Temperature

  Name	Access Path	Functionality	Unit
  iGPU Temperature	/sys/devices/platform/22100000.pcie/pci0000:01/0000:01:00.0/gpu_temperature_max	Returns a signed fractional value, e.g., -23.6, as maximum detected temperature across all therm zones on iGPU, including GR3D and NVENC/NVDEC/OFA engines.	Celsius

  • Suggested Query Method

    • 1s query interval.

    • Print engine temperature to user interface every 1s or larger interval.

  • Query interval must be greater than or equal to 10ms and greater than thermal throttling detection period. This is because the control knob is not queried more often than for thermal throttling purposes and considers message round-trip time.

• iGPU Engine Clock Frequency

  Name	Access Path	Functionality	Unit
  GPC Engine Clock	/sys/kernel/debug/gpu.0/clk/gpu_gpc/pto_counter	Returns an integer value as GR3D engine clock frequency from the same NAFLL clock source for 2 GPCs.	Hz
  NVD Engine Clock	/sys/kernel/debug/gpu.0/clk/gpu_nvd/pto_counter	Returns an integer value as video engine clock frequency driving all NVENC/NVDEC/OFA engines.	Hz

  • Suggested Query Method

    • 1s query interval. This is to align with iGPU temperature query interval because in an automotive product the iGPU engine clock frequency is not expected to change unless temperature changes.

    • Print each engine clock frequency to user interface every 1s or larger interval.

  • Query interval must be greater than or equal to 10ms or be the same as iGPU temperature query interval.