Tesla shows an AI chip roadmap with new 'SI' names and a returning Dojo

A Tesla presentation relabels its inference chips SI4, SI5 and SI6, for 'super intelligence'. Tom's Hardware reports that SI6 is targeted at 80 times the compute of SI4, and that a Dojo D3 training chip moves to chiplets with stacked HBM, after Musk cancelled the earlier wafer-scale Dojo in August 2025.

Why it matters: It suggests Tesla is again building its own training silicon, though the performance figures are targets, not measured results.

SI5 vs SI4

40x compute (Tesla target)

SI6 vs SI4

80x compute (Tesla target)

Dojo D3 memory

Stacked HBM; capacity not disclosed

List price

Not disclosed