~/TESLA/tesla-updates-ai-accelerator-roadmap-with-si6-and-space-ready-dojo-d3

Tesla Updates AI Accelerator Roadmap with SI6 and Space-Ready Dojo D3

Tesla has updated its AI accelerator roadmap, revealing that its upcoming SI6 inference chip will deliver 80 times the computing power of HW4, while the Dojo D3 training accelerator will transition to a multi-die design with stacked HBM memory and space deployment capabilities. This roadmap highlights Tesla's continuous push into custom silicon design to power its autonomous driving and data center workloads independently. Furthermore, optimizing chips for space deployment signals a potential expansion of Tesla's hardware applications into aerospace environments. The SI6 inference chip is expected to reach 17,280 to 46,080 TOPS and feature 144GB of memory. Meanwhile, the Dojo D3 training accelerator abandons the previous wafer-scale SRAM architecture in favor of a 2.5D/3D packaged multi-die layout with high-bandwidth memory (HBM).

## BACKGROUND

Tesla's hardware lineup includes inference chips used inside vehicles for Full Self-Driving (FSD) and training chips like Dojo designed for large-scale data center machine learning. Inference involves executing a trained neural network in real time, whereas training requires massive computational power to build those neural networks from data.

## REFERENCES

## KEYWORDS

#Tesla#AI Accelerators#Hardware Roadmap#Custom Silicon#Autonomous Driving

$ subscribe --daily