▸case-01 I have a PyTorch YOLOv8 object detection model ready for deployment on a mobile robot powered by an NVIDIA Jetson Orin. I want to achieve the lowest inference latency possible for single-frame camera inputs. Should I deploy the PyTorch `.pt` file directly using TorchScript on the device CPU, or how should the model be compiled for maximum hardware acceleration? | pass→pass | 14,027 | 32,055 | +129% | 1 | 1 | 0% | 2,499 | 3,578 | +43% | 0 | 0 | — |
▸case-02 When optimizing a convolutional neural network for real-time edge execution on Jetson hardware, I plan to apply INT8 quantization. Should I perform simple static quantization without any representative data, or what specific mechanism is required during engine creation to maintain detection accuracy? | pass→pass | 14,461 | 16,757 | +16% | 1 | 1 | 0% | 2,416 | 3,542 | +47% | 0 | 0 | — |
▸case-03 We are building an edge vision platform on Jetson AGX Orin that processes four concurrent 4K H.265 video streams for obstacle detection. Our developer suggested decoding frames with Python OpenCV `cv2.VideoCapture` and running PyTorch in separate threads. What architecture should be used instead for zero-copy hardware acceleration across video decoding and inference? | pass→pass | 27,699 | 22,157 | -20% | 1 | 1 | 0% | 4,158 | 4,004 | -4% | 0 | 0 | — |
▸case-04 On an NVIDIA Jetson Orin hardware platform with dual Deep Learning Accelerators (DLA) alongside the primary GPU, we want to run a lightweight classification model continuously without consuming GPU CUDA cores needed for point cloud processing. How should the execution target be configured? | fail→pass | 13,846 | 15,586 | +13% | 1 | 1 | 0% | 2,372 | 2,760 | +16% | 0 | 0 | — |
▸case-05 In a real-time robot control loop, our single-batch inference call on the GPU exhibits high per-frame CPU overhead due to repeated kernel launch enqueue calls. How can we record and replay the GPU kernel execution sequence to eliminate host CPU overhead? | pass→pass | 21,520 | 19,408 | -10% | 1 | 1 | 0% | 3,017 | 3,592 | +19% | 0 | 0 | — |
▸case-06 We need to pass camera frames from a ROS2 image topic into an edge GPU model runner with minimum memory latency. The team currently uses `cv_bridge` to copy frames into CPU NumPy arrays before transferring them to GPU memory via `cudaMemcpy`. What approach eliminates this CPU copy bottleneck? | pass→pass | 17,847 | 20,216 | +13% | 1 | 1 | 0% | 2,966 | 3,817 | +29% | 0 | 0 | — |
▸case-07 When processing 8 camera feeds simultaneously on an edge hardware module, what component in a GStreamer-based NVIDIA pipeline aggregates individual frame buffers into a unified batch tensor before feeding the neural network? | pass→pass | 6,001 | 7,848 | +31% | 1 | 1 | 0% | 1,034 | 1,660 | +61% | 0 | 0 | — |
▸case-08 We want to export a custom PyTorch segmentation model for deployment on an embedded NVIDIA system. Should we compile directly from PyTorch source code on the edge device during application startup, or what standardized intermediate format should be exported first? | fail→pass | 16,014 | 16,017 | +0% | 1 | 1 | 0% | 2,760 | 3,406 | +23% | 0 | 0 | — |
▸case-09 Our deep neural network is too large to meet the strict 15 millisecond latency budget on a Jetson Nano even after TensorRT FP16 quantization. What architectural reduction techniques should be applied before hardware engine building? | pass→pass | 17,878 | 19,549 | +9% | 1 | 1 | 0% | 2,865 | 3,638 | +27% | 0 | 0 | — |
▸case-10 We need to measure component-level latency, memory bandwidth, and power consumption of our neural network execution on NVIDIA Jetson hardware. What official NVIDIA command-line and GUI profiling tools should be used instead of standard Python `time.time()`? | pass→pass | 21,396 | 19,725 | -8% | 1 | 1 | 0% | 3,771 | 3,581 | -5% | 0 | 0 | — |
▸case-11 We are deploying an ONNX model in a C++ application on a Jetson Orin device using ONNX Runtime. By default, ONNX Runtime uses the CPU Execution Provider. Which specific execution provider should be configured to leverage the Jetson GPU hardware acceleration engine? | pass→pass | 10,044 | 7,158 | -29% | 1 | 1 | 0% | 1,482 | 1,526 | +3% | 0 | 0 | — |
▸case-12 During INT8 compilation of a custom topology, several custom layers fail to build because they lack INT8 quantization kernels in TensorRT. How should the builder flags be configured so that unsupported layers run in FP16 while supported layers run in INT8? | pass→pass | 13,966 | 18,535 | +33% | 1 | 1 | 0% | 2,556 | 3,265 | +28% | 0 | 0 | — |
▸case-13 We have trained a Reinforcement Learning actor network in PyTorch for high-frequency (1 kHz) joint torque control on a quadrupeds robot. How should the neural network policy be packaged for low-latency microsecond execution in a ROS2 C++ control node? | fail→fail | 24,776 | 33,658 | +36% | 1 | 1 | 0% | 4,563 | 6,584 | +44% | 0 | 0 | — |
▸case-14 Before performing field testing and validation of an autonomous rover on NVIDIA Jetson Orin hardware, how should the system power modes and clock frequencies be configured to ensure maximum computational performance during latency benchmarks? | pass→pass | 13,857 | 14,087 | +2% | 1 | 1 | 0% | 2,306 | 2,630 | +14% | 0 | 0 | — |
▸case-15 In an object detection pipeline on Jetson, Non-Maximum Suppression (NMS) in PyTorch CPU code takes 8ms out of the total 12ms frame budget. How can we eliminate this post-processing bottleneck? | pass→pass | 19,265 | 20,230 | +5% | 1 | 1 | 0% | 2,842 | 3,779 | +33% | 0 | 0 | — |
▸case-16 In a video analytics pipeline, we need to detect vehicle bounding boxes and then run a secondary classification model on each detected vehicle. How does DeepStream handle this two-stage detection and classification pattern natively in GStreamer? | pass→pass | 15,358 | 20,859 | +36% | 1 | 1 | 0% | 2,697 | 3,442 | +28% | 0 | 0 | — |
▸case-17 Our vision model on Jetson must accept camera streams with dynamic batch sizes (ranging from 1 to 4 frames depending on active cameras). How should TensorRT optimization profiles be configured during engine build? | pass→pass | 14,726 | 15,581 | +6% | 1 | 1 | 0% | 2,565 | 3,026 | +18% | 0 | 0 | — |
▸case-18 When distributing trained models to a fleet of 50 Jetson AGX Orin robots running identical hardware and JetPack versions, what pre-built artifact should be generated and distributed to avoid on-robot model compilation at boot time? | pass→pass | 8,002 | 17,003 | +112% | 1 | 1 | 0% | 1,167 | 2,709 | +132% | 0 | 0 | — |
▸case-19 When receiving 16 IP camera H.264 streams on an NVIDIA Jetson Xavier, high CPU utilization during video decoding is starving the ROS2 motion planning threads. Which hardware-accelerated video decoder element should be used in the GStreamer pipeline to offload decoding from CPU to NVDEC? | pass→pass | 9,803 | 8,995 | -8% | 1 | 1 | 0% | 1,639 | 1,819 | +11% | 0 | 0 | — |
▸case-20 In a multi-stream edge inference system handling 4 camera streams, frame jitter occurs because inference calls are executed synchronously in the main processing thread. How should the pipeline buffer management and asynchronous execution be structured? | fail→fail | 22,715 | 19,481 | -14% | 1 | 1 | 0% | 3,104 | 3,758 | +21% | 0 | 0 | — |
▸case-21 Our team compiled a TensorRT engine file on a host x86 workstation with NVIDIA RTX 3090 GPU and wants to copy the `.plan` file directly to a Jetson Orin device running JetPack. Will this binary engine run directly without issue? | pass→pass | 9,345 | 14,481 | +55% | 1 | 1 | 0% | 1,687 | 2,854 | +69% | 0 | 0 | — |
▸case-22 We are designing a cloud-based cluster deployment on AWS to serve a 70B parameter Large Language Model to web applications using vLLM on a fleet of A100 GPUs. How should we configure Kubernetes HPA auto-scaling based on queue depth and GPU memory metrics? | pass→pass | 22,344 | 28,824 | +29% | 1 | 1 | 0% | 3,739 | 4,780 | +28% | 0 | 0 | — |
▸case-23 We need to deploy an interactive image background removal tool inside a user's web browser (React single-page application) so user photos never leave their device. What web runtime technologies should be used to run the model client-side in the browser? | pass→pass | 15,958 | 17,768 | +11% | 1 | 1 | 0% | 2,741 | 3,397 | +24% | 0 | 0 | — |
▸case-24 We are writing a custom C++ CUDA kernel to accelerate a novel loss function during PyTorch training on our desktop workstation equipped with dual RTX 4090 GPUs. How should the PyTorch C++ extension be compiled using `torch.utils.cpp_extension`? | pass→pass | 16,156 | 24,797 | +53% | 1 | 1 | 0% | 3,170 | 5,303 | +67% | 0 | 0 | — |