Check out more Kit Close-up videos.
I took a quick look at NVIDIA's Jetson AGX Orin Development Kit built around NVIDIA's Jetson AGX Orin module (Fig 1). The form factor is compatible with the earlier Jetson AGX Xavier, but the Orin can deliver eight times the performance at 275 TOPS. The development kit provides a way to get started with the module.
The Jetson AGX Orin module (Fig. 2) is built around a CPU, GPU, and accelerators designed to handle machine-learning (ML) models. The CPU complex includes a dozen, 64-bit Arm Cortex-A78AE v8.2 cores that have access to a 3-MB L2 and 6-MB L3 cache. The GPU incorporates NVIDIA's new Ampere architecture. It includes 2,048 NVIDIA CUDA cores and 64 Tensor cores.
In addition, there are two NVIDIA Deep Learning Accelerators (NDLAs) v2.0 along with a PVA v2.0 vision accelerator. The module adds 32 GB of 256-bit LPDDR5 DRAM with a 204.8-GB/s bandwidth. Nonvolatile storage includes 64 GB eMMC 5.1. The M.2 socket on the kit's carrier board can be utilized to extend the on-module storage.
The smaller M.2 E-key is used for the Wi-Fi 802.11ac and Bluetooth support. There are two antennas; one is hidden in a cover that also hides the PCI Express (PCIe) socket. On top of that, a 40-pin header provides access to interfaces like I2C, CAN, and serial ports.
NVIDIA's software tools and stacks compatible with ML frameworks like PyTorch and TensorFlow are what make Jetson AGX Orin so compelling (Fig. 3).
I will be taking a closer look at these tools and stacks in a future hands-on article. In the meantime, check out the Kit Close-up video above.