Tesla Dojo

Tesla's from-scratch AI-training supercomputer, built around its own D1 processor rather than merchant GPUs. Unveiled at Tesla's first AI Day on 19 August 2021 with the D1 chip and Training Tile module, detailed further at Hot Chips 34 in August 2022, and put into production per Tesla's own Q2 2023 shareholder update (19 July 2023), independently corroborated by TechCrunch's retrospective timeline and by Digitimes' contemporaneous reporting. The full ExaPOD design combines 120 Training Tiles of 25 D1 chips each, 3,000 chips and 1,062,000 usable cores, rated at roughly 1.1 exaflops BF16/CFP8 in Tesla's original announcement and in most subsequent press coverage (Videocardz, HPCWire, Data Center Dynamics); The Next Platform's Hot Chips writeup separately cites a 20-exaflops figure for the full system that looks like an outlier against that widely repeated 1.1 EFLOP figure, so rpeak here uses 1.1 EFLOP (1,100 PFLOP) and flops_basis is vendor_peak rather than hpl_measured, since Dojo was never submitted to TOP500. The exact site was not pinned down across the sources fetched: early builds were associated with a 2022 test that tripped a local substation, while later scale-out plans (mid-2024) targeted Gigafactory Texas, so no site_id is recorded here. Tesla disbanded the Dojo team and shut the program down in August 2025 (confirmed independently by TechCrunch and Electrek), with Elon Musk calling Dojo 2 'an evolutionary dead end' in favor of the AI6 inference chip; Tesla's separate, Nvidia-GPU-based Cortex cluster (see tesla-cortex) is what absorbed the training workload. A Dojo 3 revival was reported in early 2026 but is not yet documented well enough by independent sources to add as its own row.