Four years ago, we first mentioned the Open Compute project, through which Facebook opened the design of its main servers. Thanks to Open Compute, any company can adopt those configurations and use them in their own environments, but now another element is added. Its code name is Big Sur, and it is the server that Facebook uses to train artificial intelligence software. As of today, Big Sur is open source.
The growing demand for open source AI
More and more companies are interested in machine learning and artificial intelligence. The amount of data to process continues to increase, as does its general complexity, so systems and algorithms must be up to the task. For some of the big names on the Web, the best way to stimulate the development of both fields is with open source technology. In the early days of November, Google opened its TensorFlow platform, so that experts and enthusiasts alike can experiment with its code. Obviously, all software needs a piece of hardware to run it, and with that in mind, we arrive at Big Sur.
Big Sur in detail
Big Sur is the latest server design from Facebook, specially prepared for large-scale artificial intelligence processes. In other words, it is the hardware set on which the social network “trains” its programs, responsible for carrying out tasks such as image recognition (an element associated with the famous automatic tagging). The Big Sur design allows the inclusion of up to eight GPUs per Open Rack unit, with a maximum TDP of 300 watts per chip. As expected, Big Sur is based on Nvidia's Tesla Accelerated Platform, and according to the official publication, this design can train programs twice as fast, as well as explore networks twice as large as the previous generation.
Now, the fact of being based on Nvidia hardware does not mean that options are limited to that company. Facebook recognizes that Big Sur was imagined around the Tesla M40 cards, but those interested in this design will be able to implement a wide variety of PCI Express cards, a detail that will surely allow precisely adjusting the cost-performance ratio. The server offers very easy access to components, to such an extent that the motherboard can be replaced in a minute if necessary. Facebook has not yet indicated the exact date on which it will open the Big Sur design, but the next Open Compute conference will be in March, and we calculate that it will wait until then.
