Computing the Impossible #TechVision F@h demonstrates the potential power of going to get them the performance they needed parallel computing, and what can be achieved to train their models – no matter how many when sheer scale is combined with the power were clustered in a machine. So Tesla designed of today’s CPUs and GPUs. But F@h is a bit of the D1 Dojo chip that was built specifically to an outlier because of its distributed nature. The run the computer vision neural networks that most powerful supercomputer clusters today are underpin the company’s self-driving technology. localized, and almost always commissioned to The company designed a unique architecture run small, highly specific projects. For instance, to cluster these together, and thus the Dojo the Sierra supercomputer at Lawrence Livermore supercomputer was born. National Laboratory runs nuclear weapons test simulations in place of setting bombs off in But here’s the thing: While the confluence 176 real life. However, the demands for cutting of access to a unique data set and company edge computing power that were once isolated ambition led Tesla to this very specific need, to low-frequency events are now becoming the problem they are trying to solve isn’t niche increasingly necessary to the regular operations at all – in fact it’s at the heart of the industry. of a post-digital business. That’s why Tesla It’s easy today to think of Tesla as an anomaly. decided to build their own. But what happens when their cars start driving themselves? What happens when they can In June 2021, Tesla unveiled Dojo, a provide a feature no other car manufacturer can supercomputer designed entirely in-house and deliver, because those competitors simply don’t built for one purpose: making self-driving cars a have the technology to design it? Companies 177 reality. Tesla has long been at the forefront of need to ask themselves: Is this what’s required technology leadership, and this is an example to compete in tomorrow’s market? Are industry where they pushed technology beyond what – or company – specific chip designs and they could get from specialty vendors. For years, architectures the white space where they can Tesla had been gathering massive amounts of find a source of competitive advantage? driving data from its cars, and it was clear no amount of off-the-shelf generalized chips were Introduction // WebMe // Programmable World // The Unreal // Computing the Impossible 79