Big Sur is built specifically to train neural networks how to respond to inputs. After initially building servers using off-the-shelf standard industry components, Facebook realized its advanced software was being held back by its hardware. It has built a custom GPU-based solution, Big Sur, which is capable of hosting its AI software.
Big Sur uses the NVIDIA Tesla Accelerated Computing Platform to distribute load across eight graphics processors, each drawing up to 300 watts of power. This advance means Facebook can train its neural systems to be twice as large as previous generation models twice as fast.
The system isn’t just about raw compute power though. Big Sur gives administrators and technicians easy access to its components for maintenance and upgrades. The entire motherboard assembly can be removed in under a minute, without any special training besides basic knowledge of a computer’s internals.
Big Sur will be used to help Facebook develop more complex software using artificial intelligence and machine learning. The company has already developed projects that can play games on their own, provide meaningful answers to questions and be taught how to do something new.
With the open-sourcing of Big Sur’s design comes the possibility of other people being able to create similar software. Building hardware specifically for AI projects can be as difficult as creating the project itself. Facebook’s solution can be used by teams of engineers as a starting point for something capable of running their software.
Facebook said:
“We want to make it a lot easier for AI researchers to share techniques and technologies. As with all hardware systems that are released into the open, it’s our hope that others will be able to work with us to improve it. We believe that this open collaboration helps foster innovation for future designs, pushing us all one step closer to building complex AI systems that bring this kind of innovative to our users and, ultimately, help us to build a more open and connected world.”
Facebook has already open-sourced much of the Torch software side of its deep learning systems. By revealing the inner workings of its current-generation hardware, the company is sharing everything it has built around AI with the world, giving more people the chance to start developing intelligent systems.
Facebook hasn’t said when the plans for Big Sur will be fully open-sourced. It will submit all the relevant design materials to the Open Compute Project, a body that oversees the sharing of datacentre infrastructure designs between companies, to begin sharing with the community.
