NVIDIA Announces A2 Entry-Level Edge Tensor Core GPU Based on Ampere Architecture

NVIDIA has today launched the latest addition to its data-centric graphics card lineup, called A2 Tensor Core GPU. The company envisions the A2 Tensor Core GPU as an entry point of accelerators that power machine learning and artificial intelligence at the edge. It is based on Ampere architecture and built on Samsung’s 8 nm node. The GPU has GA107 GPU with 1280 CUDA cores enabled, where the full GA107 has 3072 cores. This GPU has only 1280 CUDA cores because it is configured as a low-power, low-profile body with a configurable TBP (total board power) of 40-60 Watts, depending on the customer use case.

It directly replaced the previous generation T4 accelerator, an edge device, and promises to bring 20-30% higher performance in the same power envelope. As far as memory, it has 16 GB of GDDR6 memory running on a 128-bit bus and 200 GB/s memory bandwidth. The GPU core clocks at 1440 MHz base and 1770 MHz boost frequency, which outputs 4.5 TeraFLOPs of FP32 compute. A2 represents a significant upgrade over the previous T4 because it offers a 60% better price to performance ratio and 10% better power efficiency than the last generation. Overall, this release represents an excellent upgrade for customers that need basic AI processing at the edge, and no, gaming is not supported on these as video output is disabled.