Tesla’s Dojo on Wafer Processor Moves into Production
Tesla’s Dojo processor has moved on to mass production and will be deployed soon. Dojo is a system on a wafer processor that consists of a 5-by-5 series of individual processor chips that are fixed on another wafer called a carrier wafer and are interconnected by Taiwan Semiconductor’s InFo technology for wafer interconnections. Sounds complex? Let’s simplify it.
Tesla’s Dojo will be deployed soon
Tesla Dojo Training Tile (the company calls it a training tile) is actually a pack of 25 powerful processors that are joined together to act as one ultra-powerful processor, utilizing TSMC’s connectivity technology, which is specifically designed for high-speed connectivity. Now it has started making a little sense, but first a bit of background on why Tesla needed its own processor when semiconductor giants like Nvidia, IBM, and Intel have already been making hardware for computers and AI.
Tesla’s decision to enter AI, chip design, and supercomputing shows how the company, and especially Elon Musk, sees innovation as more than just corporate business expansion. Tesla is trying to develop cars with self-driving capabilities from the start, and all the cars it has sold transmit data to Tesla to help the company build the system. So Dojo is also a part of its vertical integration.
Dojo is quite different from other available AI accelerators and tools for super computing. For example, Nvidia’s A100 or A200 GPUs are embedded in many super computers around the world, but these and the others are mostly designed to cater to a wide range of tasks, including research, big data processing, complex simulations, and many more. But Dojo is designed for a specific task like AI computer vision driven by real-world data; what else Dojo will be capable of is unknown, as the company has not revealed much.
Wafer-based systems are more efficient
As said above, Dojo consists of 25 individual high-performance processors, so it has a ravenous appetite for power and requires an effective and sophisticated cooling system. According to Tom’s hardware, to feed the power, Tesla uses a voltage-regulating system, which in itself is a complex module that delivers 18,000 amps of power to the processor. With these high levels of power, the system dissipates 15,000 watts of heat, and to work efficiently, it requires a liquid cooling system.
Wafer-based accelerators, such as Tesla’s Dojo, are substantially more effective and give better performance than other multiprocessor systems. They benefit from low latency between the communication cores, high bandwidth, and better energy efficiency due to the system’s wafer technology.
Only Tesla and Cerebras have systems on wafer designs up until now, but other industry players are also expected to get in because of the higher efficiency and lower latency. However, there are some limitations to these accelerators as well; wafer systems currently have to rely on on-chip memory, which is usually fixed and may be insufficient for different types of use cases. But it is expected that this will be overcome when the next generation of on the wafer platform enables 3D stacking of HBM4 memory chips on accelerator tiles.
