AKHET A.I

Looking at the hardware and requirements for data handling of AI applications, such as inferencing or deep learning (in the so-called data center), one finds that each core component of the computers must not fall below certain performance values. For example, single or multi-socket CPUs based on Intel® Xeon® or AMD EPYC™ processors are recommended by NVIDIA. For the management of I/O and other workloads, a lower limit of 6 CPU cores is drawn here. RAM distributed over all CPU sockets and memory channels brings optimum performance here if the quantity is at least 1.5 times larger than the GPU frame buffer memory used. At least one PCIe slot x16 is required for each GPU, at least with the bandwidth of Generation 3.0. Network adapters are required for networking AI systems, for processing the large data volumes or for HPC tasks, which on the one hand offer bandwidths of up to 200 Gb/s Ethernet, and on the other hand reduce the overhead of the CPU to zero in the process. NVMe-based “drives” are to be used for storage. One per CPU socket so that the optimum application performance can be accessed.

The topic of artificial intelligence (AI) is already more than 65 years old, and global market intelligence leaders in 2020 predicted its strong penetration in enterprises by 2024.

The reason for this is the interplay between research and development of machine learning (ML) algorithms or frameworks, the availability of powerful hardware capable of analyzing huge amounts of data in real time, and of course the need for and ever-increasing interest in AI applications.

The importance of this technology for society and science is shown in this diagram: Current surveys and the evaluation of public documents and patents resulted in a top 10 list of the most important technology trends of the next few years. Thus, within this top 10 list, Artificial Intelligence is also mentioned. But unlike all other trends, AI is listed three times.