Cirrascale Artificial Intelligence Cloud: Experiments with AI and Deep Learning

Over the past few years, artificial intelligence has become the talk of the town. It is both about new developments in the field of weak AI, and about the opinion of experts that artificial intelligence can be dangerous for humanity. But be that as it may, specialists gradually improve the weak form of AI, making it more efficient and productive. For example, IBM is gradually introducing the Watson cognitive system into many of its products, as well as the development of third-party companies. Google is ahead of the rest of the world with AlphaGo, and the division of this corporation DeepMind creates new forms of AI.

An important role in such projects is played by machine learning. This area is also developing very quickly, gradually appearing in various directions of development of science, technology, and human life. RoboMobi, online marketing, cybersecurity, financial transactions, military. This is only a small fraction of the areas where machine learning is used, without which AI is impossible. The hardware that is involved here must be very powerful. Not every developer can afford to purchase the necessary equipment. And here comes a specialized data center.

In California, Cirrascale appeared recently, which built a data center that is designed for AI and machine learning. Cirrascale is a subsidiary of Poway, a company that supplies high-performance equipment. In addition, Poway is a cloud service provider. Now Cirrascale is engaged in designing an infrastructure for deep learning. The data center of the company ensures the operation of the cloud service provided in the form of SaaS (software as a service). The principle of operation is similar to Amazon Web Services, although there are significant differences.
')
Firstly, virtual instances are not used here, but only powerful servers with a fairly high energy consumption rate. The client of the company has at its disposal such a server (or several) to run specialized software. This was done to ensure that clients were engaged only in the development of the software part of their projects. "Iron" is provided by the data center. The fact is that not every developer is able to assemble an HPC cluster, ensuring its uninterrupted work.

In the data center Cirrascale, as already mentioned, the hardware consumes a lot of energy. One rack here has about 30 kW, while in the average data center the rack has from 3 to 5 kW, rarely - about 10.

It can be assumed that powerful equipment will generate a lot of heat. And indeed it is. To cool the servers, a proprietary liquid cooling system developed by ScaleMatrix is ​​used here. In addition, an air cooling system is used. Only here the air is not along the racks, but from the bottom up, with very high speed. Each server rack has its own microclimate with a water cooling system and air circulation system. All this allows you to be sure that the racks do not affect the thermal regime of each other.



As for the servers, they also differ from the servers of ordinary data centers. Here the main role is played by the GPU, in each server there is a whole group of them. Basically, this is Tesla GPU from Nvidia, which work together with Intel Xeon processors. The most powerful cloud system in the data center is a server with 16 GPUs of the specified type. The cost of it, of course, rather big. Monthly rent will cost $ 7,500.

GPUs in one system are connected in a special way with each other, so that all elements are a single whole. This ensures maximum throughput while helping to maintain high performance and scaling.


The location of individual elements in the supercomputer DGX-1

The Nvidia DGX-1 supercomputer has a similar configuration. It was designed specifically for deep learning, with the interconnection of the GPU using NVLink technology.

Now they write and say a lot about deep learning, Google, Facebook and several other companies are actively implementing developments from this sphere. But most of the companies are only at the beginning of the path to the introduction of deep learning and AI. The reason was stated above - not everyone has the equipment to implement their theoretical developments.

Perhaps now there will be more projects on the topic of deep learning, because, as we see, companies are emerging that supply hardware for software components. Results are already appearing. For example, one of the Cirrascale clients creates a weak form of AI for driving. Moreover, to assess the situation, this car uses not separate images, but video. Without powerful equipment, working with such a system would be difficult.

Representatives of this company are confident that connecting the GPU to each other within the server is a technology that will become the standard in the near future as neural networks evolve. Perhaps over time, the GPU will be replaced by something else, but so far it is far from it. But the providers of services and equipment to scientists and companies that work in the field of neural networks, in-depth training, AI, will become more and more. By the way, in the process of writing a post, it turned out that Google now provides a similar service.