Facebook releases system design for deep learning servers
Facebook made the technology of its Big Sur servers public. The systems are optimized for deep learning calculations to allow computers to recognize speech and images.
Big Sur is compatible with the Open Rack V2 standard from the Open Compute Project. This project was set up by Facebook to make as much documentation about data center systems public as possible, and Big Sur’s design and specifications are also distributed via Open Compute. The specific details of the system have yet to be made public.
Facebook has developed Big Sur together with Nvidia and by default the system contains eight Tesla M40 cards. However, according to the company, Big Sur can also be equipped with other PCI-e cards. Despite the considerable amount of GPU computing power, the server would not require special cooling. The design was built with easy access to hardware components in mind. This way, components that need to be replaced regularly, such as hard disks and dimm memory, can be replaced quickly.
According to the social network, the motherboard can also be replaced within a minute, where that would take an hour with previous comparable servers. Moreover, that replacement could all be done toolless; only a screwdriver is needed to exchange the cpu cooler. Green indicates where technicians can touch the parts.
According to Facebook, Big Sur is twice as fast as the previous generation of deep learning systems the company used. Machine learning and artificial intelligence are becoming increasingly important for tech giants such as Facebook, Google, Microsoft and IBM.
The techniques enable the companies to build services around ‘understanding’ of large complex data sets. For example, recognizing voices and images, choosing content for news feeds, personal assistants and translating content rely on the techniques. By making the technology public, the companies hope to boost innovation, which they themselves can benefit from. For example, Google made its self-learning software TensorFlow open source at the beginning of November.