Tesla’s AI chip roadmap promises bigger gains, but the dates are years away

Tesla has outlined a roadmap for future AI accelerators, including a sixth-generation inference chip it expects to deliver 80 times the compute performance of today’s benchmark. The same roadmap brings back Dojo, now as a chiplet-based training design using stacked high-bandwidth memory, rather than the earlier wafer-scale approach.

Watch Desk analysis

What happened

Tom’s Hardware reports that Tesla presented the roadmap on 4 November. The report says the sixth-generation inference accelerator, SI6, is expected in 2029 or 2030. Tesla’s projected 80-fold performance improvement is measured against its current-generation chip; the article’s estimates of absolute performance depend on assumptions about that baseline.

The roadmap also describes SI5, an earlier multi-chip-module design expected to deliver a 40-fold improvement over the same current benchmark. For training, Dojo D3 is described as combining chiplets, 2.5D and 3D packaging, and stacked HBM. Tesla has not disclosed D3’s performance, memory capacity, manufacturing process or development schedule, according to the report.

Why it matters

This is a substantial, if distant, computing plan: inference chips for Tesla’s vehicles and a reworked accelerator for training AI models. The proposed memory and packaging choices matter because they shape what these chips can handle and how difficult they may be to manufacture.

But a roadmap is not a delivered processor. The big performance figures are projections, and the report notes that comparisons are complicated by uncertainty about the current chip’s absolute performance. The practical test will be what Tesla ships, when it ships it, and how those parts perform outside a slide.

Our read

The interesting detail is not simply a very large multiplier. It is Tesla pursuing two different hardware paths: inference accelerators with multi-chip packaging, and a redesigned Dojo training chip with stacked HBM. That gives the roadmap more substance than a renamed product family, while leaving plenty to prove before the numbers earn their headlines.

What to watch

• Whether Tesla publishes more detail on SI5 and SI6 specifications, power use and schedules.
• Whether Dojo D3 advances beyond its broad design description.
• How future performance claims compare with independently tested hardware.

Discussion spark: For an AI hardware roadmap, what should count as convincing evidence: ambitious performance targets, a credible manufacturing plan, or working chips in customers’ hands?

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.