NVIDIA A2, here is the new inference accelerator based on Ampere GA107 GPU

Is called A2, or in a longer version “A2 Tensor Core” the new acceleratore entry level for datacenter presented by NVIDIA at the GTC 2021 which was held in the past few hours. It is particularly aimed at related calculations to inference and aims to replace the previous NVIDIA T4 based on the Turing architecture.

According to the US company, the NVIDIA A2 Tensor Core accelerator offers up to 20 times higher performance with regards to inference versus CPU-only servers e up to 30% higher than NVIDIA T4. At the same time it guarantees a 60% higher price – performance ratio and 10% better energy efficiency than the T4 model.

It is a low-end solution because it is based on one GPU Ampere GA107 con 1280 CUDA core and 40 Tensor cores flanked by 16GB of GDDR6 memory on 128-bit bus for a bandwidth of 200 GB / s. Basically the GPU operates at 1440 MHz and reaches 1770 MHz in boost.

Able to offer one computing power of 4.5 TFLOPs with single precision FP32 calculations, NVIDIA A2 has a TDP at 40-60 Watts and requires no external power supply. Like other server products, this solution is also passively cooled. More information on NVIDIA’s new accelerator can be found in This Page.