NVIDIA Tegra X1 (TX1) based client-side neural network acceleration is a strong complement to server-side deep learning using NVIDIA Kepler based Tegra server accelerators. In the software world, “function overloading” refers to using one procedure name to invoke different behavior depending on the context of the procedure call. NVIDIA borrowed that concept for their graphics processing pipelines. They essentially are “overloading” their graphics architecture to be a capable neural network processing architecture. Moor Insights & Strategy evaluates the Nvidia Tegra X1 in the context of neural network acceleration.
- Floating Half Precision (versus ARM, Imagination Technologies, AMD)
- Fused Operations
- How is NVIDIA’s Approach Different?
- Usage Model: Internet of Things (IoT) Comes to Car and Driver (DRIVE PX)
- Maybe You Can Drive My Car
- Figure 1: A Neural Network “Perceptron”
- Figure 2: A Car Full of Video Cameras and Their Overlapping Fields of View
- Figure 3: Not an Average Street Scene with Recognized Objects Labeled
- Figure 4: High Level View of Car and Datacenter Working Together
- Figure 5: Functional View of NVIDIA DRIVE PX Communicating with Datacenter
Click here to download the paper.
- Imagination Technologies