NVIDIA announced the launch of an entry-level accelerator card called NVIDIA A2, adding more products to its Ampere architecture lineup for data centers. NVIDIA A2 uses a single-slot, half-height, and half-length specification. It seems to use a GA107 GPU, but it has been castrated significantly, with only 1280 CUDA cores. At the same time, it is equipped with 16GB of GDDR6 memory, the bit width is 128 bits, the rate is 12.5 Gbps, and the bandwidth is 200 GB/sec. Its TDP is 40W to 60W, does not require an external power supply, adopts a passive heat dissipation structure, is not equipped with a fan, and is not equipped with a video output interface.
The basic frequency of NVIDIA A2 is 1440 MHz, and the acceleration frequency is 1770 MHz. It can provide 4.5 TFLOPS of single-precision computing power, which is lower than the GeForce RTX 3050 Ti mobile version (7.1 TFLOP) equipped with 2048 CUDA cores. Compared with the previous entry-level product T4, in the edge computing scenario, the performance has been improved by 20% to 30%, the power consumption has been reduced by 10%, and the cost performance has been improved by 60%.
Since there is no video output interface, it cannot be used for games if it is not used in a virtual machine environment. The single-slot, half-height, and half-length specifications are designed to be used in servers that value cost-effectiveness and require high compatibility and are optimized for AI inference workloads. Nvidia did not disclose the pricing of NVIDIA A2, but confirmed that it has begun supplying to OEM partners.