Benchmark, GPUExternal
Leveraging NVIDIA data-center GPUs for AI inferencing
Benchmarking the NVIDIA A2, A10, A40, A100, L4, L40, and H100 across a real production video-inference pipeline (PeopleNet, FaceDetect, face-mask classification). Practical findings on tracker choice, interval values, and where decoders become the true bottleneck.
// Key takeawayL4 delivers roughly 2x the inference throughput of T4 across every tracker tested, but only because the encoder/decoder bottleneck shifts before raw compute does.
Originally on Medium, 2023. Adrian Apap, co-author.
Read on Medium →Benchmark, EdgeExternal
Is the new NVIDIA Jetson AGX Orin really a game changer? We benchmarked it.
An early hands-on benchmark of NVIDIA's then-new Jetson AGX Orin Developer Kit against the previous-generation Jetson AGX Xavier, across inception_v4, VGG19, super-resolution, U-Net segmentation, pose estimation, YOLOv3-Tiny, and ResNet50. Run on early-access software using NVIDIA's jetson_benchmarks scripts, modified to support Orin.
// Key takeawayEven on early-access software, Orin landed at roughly 2x to 4x the inference FPS of the Xavier across every model tested, at the same form factor and price band. A step-change in what fits at the edge, not an incremental refresh.
Originally on Medium, April 2022. Adrian Apap, benchmarking team contributor.
Read on Medium →Conference TalkExternal
How to start a business on NVIDIA Omniverse
Natalia spoke at NVIDIA GTC 2023 on the practical side of building businesses on top of Omniverse, covering simulation, digital twins, and the productisation gap between Omniverse demos and real customer deployments. Joint session with NVIDIA, Reallusion, Evolver, Aireal, ipolog, and SyncTwin.
// Key takeawayThe interesting Omniverse work is not in the rendering. It is in the data plumbing between the simulation, the real-world sensors, and the operational system that consumes the output.
NVIDIA GTC Spring 2023, Session S52058. Natalia Mallia, speaker.
View LinkedIn post →Digital Twin, OmniverseExternal
Seeing double: a digital twin with a simulative spin
A walkthrough of building a real-time digital twin of an Antwerp road intersection in NVIDIA Omniverse, with weather profiles and a day-night cycle driven by live atmospheric data fed in from a deployed sensor box. Covers the 3D asset pipeline, the custom connector that pulls sensor data into the simulation, and the trade-offs between geometric fidelity and runtime cost.
// Key takeawayVisual fidelity is the easy half of a digital twin. The hard half is the connector that gets real sensor readings into the simulation fast enough that the twin reacts to actual conditions, not yesterday's data.
Originally on Medium, May 2022. Natalia Mallia, co-author.
Read on Medium →