During GTC 2021, the green giant presented the NVIDIA A2, a new low-power graphics accelerator that surprised us a lot, since it uses the GA107 graphics core, a solution that, until now, we have only seen in the GeForce RTX 3050 and RTX 3050 Ti for laptops.
The GA107 GPU that the NVIDIA A2 mounts comes configured with 1,280 shaders, which equates to a total of 10 active SM drives. The GA107 chip has, in its full configuration, 20 active SM units and 2,560 shaders, so as we can see the cut that the version used in the NVIDIA A2 has received has been important.
Obviously, this has affected the rest of its specifications. The NVIDIA A2 has 40 core tensioner To accelerate workloads associated with artificial intelligence, it has a 128-bit bus and has 16 GB of GDDR6 memory, a very generous amount considering that we are talking about a low-power solution. In total, it can reach a bandwidth of 200 GB / s, and its GPU works at a maximum of 1,770 MHz.
Taking a look at the raw power offered by this graphic throttle, we find a peak of 4.5 TFLOPs in FP32. Not bad at all, especially considering that we are talking about a graphics accelerator that has a TBP (total consumption of the plate) of between 40 and 60 watts. This means that it does not use any type of additional connector, that it reaches very low operating temperatures and that it is therefore easy to integrate into almost any professional setup.
In TF32, thanks to its tensor cores, it can easily reach a power of 9 TFLOPs, while in FP16 it reaches 18 TFLOPs. If we look at the values in INT8 and INT4, we see that it has peaks of 36 TOPs and 72 TOPs, respectively. We do not have details on the sale price of this new graphics accelerator, but NVIDIA has confirmed that it is available now through major OEMs.
The NVIDIA A2 is the most basic next-gen AI graphics accelerator that NVIDIA has released so far, but also is the one with the lowest consumption, and it should be the cheapest. These keys make it a good option for different professionals, and for specific sectors.