Low performance on Drive AGX Driver

Please provide the following info (check/uncheck the boxes after creating this topic):
Software Version
DRIVE OS Linux 5.2.6
DRIVE OS Linux 5.2.6 and DriveWorks 4.0
DRIVE OS Linux 5.2.0
DRIVE OS Linux 5.2.0 and DriveWorks 3.5
NVIDIA DRIVE™ Software 10.0 (Linux)
NVIDIA DRIVE™ Software 9.0 (Linux)
other DRIVE OS version
other

Target Operating System
Linux
QNX
other

Hardware Platform
NVIDIA DRIVE™ AGX Xavier DevKit (E3550)
NVIDIA DRIVE™ AGX Pegasus DevKit (E3550)
other

SDK Manager Version
1.8.0.10363
other

Host Machine Version
native Ubuntu 18.04
other

I managed recently to install our Python/CuPy based Radar signal processing framework on Drive AGX Xavier dev kit. After doing some comparisons to other GPUs, I’m surprised that Drive AGX is significantly slower , e.g about 40% slower in comparison to my modest laptop’s GPU.

Please find attached the profiling results for the same benchmark done on a Quadro P1000 Mobile GPU as well as on Drive AGX Xavier.

Here is a snapshot of a radar cycle on Quadro P1000. Notice there are approximately no gaps between the kernels:

Here is a similar snapshot on Drive AGX Xavier - it looks to me that it takes more time to schedule kernels on the OS level than the iGPU needs for computation.

I suspect that the Nvidia Driver API on Drive AGX is taking too much time. But I would appreciate any opinions or suggestions how to improve the performance of Drive AGX .

FYI I have Python 3.8 and CuPy v9.6.0 for all benchmarks but Cuda 10.2 on Drive AGX and Cuda 11.4 for Quadro P1000.

AGXDrive.nvvp (936.6 KB)
QuadroP1000Mobile.nvvp (979.2 KB)

Dear @muhammed.al.kadi,
Could you also share cudaDeviceQuery sample output on Quadro P1000 GPU?

Dear @muhammed.al.kadi,
FYI I can see that Quadro P1000 has 1.89 FP32 Tflops where as DRIVE AGX Xavier iGPU has 1.3 FP32 TFlops. So if you use FP32 operations, perf drop is expected. We recommend to use FP16/INT8 to better results in inference.

Here is the device query on the Quadro P1000

./deviceQuery Starting…

CUDA Device Query (Runtime API) version (CUDART static linking)

Detected 1 CUDA Capable device(s)

Device 0: “Quadro P1000”
CUDA Driver Version / Runtime Version 11.4 / 11.4
CUDA Capability Major/Minor version number: 6.1
Total amount of global memory: 4032 MBytes (4227465216 bytes)
(004) Multiprocessors, (128) CUDA Cores/MP: 512 CUDA Cores
GPU Max Clock rate: 1519 MHz (1.52 GHz)
Memory Clock rate: 3004 Mhz
Memory Bus Width: 128-bit
L2 Cache Size: 524288 bytes
Maximum Texture Dimension Size (x,y,z) 1D=(131072), 2D=(131072, 65536), 3D=(16384, 16384, 16384)
Maximum Layered 1D Texture Size, (num) layers 1D=(32768), 2048 layers
Maximum Layered 2D Texture Size, (num) layers 2D=(32768, 32768), 2048 layers
Total amount of constant memory: 65536 bytes
Total amount of shared memory per block: 49152 bytes
Total shared memory per multiprocessor: 98304 bytes
Total number of registers available per block: 65536
Warp size: 32
Maximum number of threads per multiprocessor: 2048
Maximum number of threads per block: 1024
Max dimension size of a thread block (x,y,z): (1024, 1024, 64)
Max dimension size of a grid size (x,y,z): (2147483647, 65535, 65535)
Maximum memory pitch: 2147483647 bytes
Texture alignment: 512 bytes
Concurrent copy and kernel execution: Yes with 2 copy engine(s)
Run time limit on kernels: Yes
Integrated GPU sharing Host Memory: No
Support host page-locked memory mapping: Yes
Alignment requirement for Surfaces: Yes
Device has ECC support: Disabled
Device supports Unified Addressing (UVA): Yes
Device supports Managed Memory: Yes
Device supports Compute Preemption: Yes
Supports Cooperative Kernel Launch: Yes
Supports MultiDevice Co-op Kernel Launch: Yes
Device PCI Domain ID / Bus ID / location ID: 0 / 1 / 0
Compute Mode:
< Default (multiple host threads can use ::cudaSetDevice() with device simultaneously) >

deviceQuery, CUDA Driver = CUDART, CUDA Driver Version = 11.4, CUDA Runtime Version = 11.4, NumDevs = 1
Result = PASS

Actually I see that the single kernels are executed faster on Drive AGX GPU than on Quadro P1000. But the time it takes to launch them via the Cuda Driver API is remarkably longer.
I took a snapshot to show how much it takes to schedule the first 5 kernels in a radar cycle.
Here is the one of Quadro P1000:


And this is the one of Drive AGX:

Dear @muhammed.al.kadi,
But the time it takes to launch them via the Cuda Driver API is remarkably longer

If it is about CUDA kernel latency, it should be in order of micro secods (like 10-30 us). Note that, Desktop CPU is more powerful than Tegra CPU which can effect CUDA calls launch latency time.

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.