TSMC to Produce Tesla Dojo AI Training Modules with InFO_SoW

TapTechNews May 21, according to DigiTimes, TSMC announced that it has begun to produce Tesla's Dojo AI training modules using its InFO_SoW technology, with the goal of increasing computing power by 40 times by 2027 through more complex wafer-level systems.

InFO_SoW (Integrated Fan-Out Wafer-Level System-in-Package) is an improved model of the 'InFO' technology applied to high-performance computers and is also a wafer-scale super-large packaging technology, mainly including parts such as a wafer-shaped heat dissipation module (Plate), a group of silicon chips (SiliconDie), InFORDL, a power module, and a connector.

InFO_SoW arranges multiple silicon chips (SiliconDie, or Chip) horizontally on the module (similar in size to the wafer size), and then connects the chips and input/output terminals to each other through the 'InFO' structure, thus differing from the 'InFO_SoIS (System on Integrated Substrate)' technology that stacks two 'InFO's.

Generally speaking, the chip area produced by InFO_SoW technology is relatively large, and it can integrate a large-scale system (consisting of a large number of silicon chips) on a circular plate-shaped module (wafer-shaped module) with a diameter of about 300mm; and by adopting the InFO technology, it can obtain a more compact and higher-density integrated system compared to the traditional module.

In the information previously released by TSMC, InFO_SoW has more advantages than the Multi-chip-module (MCM) and 'InFO_SoW' using the FlipChip technology. TapTechNews found that compared to MCM, the width and spacing of the interconnecting wires are reduced by half, and the wire density is doubled. In addition, the data transmission speed per unit area is also doubled; the impedance (Impedance) of the power supply network (PDN) is significantly lower than that of MCM, only 3% of that of MCM.