TESLA D1 TACKLES AI TRAINING: Custom Design Combines 25 Chips Using Fanout Packaging

摘要

Elon Musk likes unorthodox approaches. For his latest surprise, Tesla has developed its own AI training chip, aiming to replace Nvidia GPUs in its massive data center. Its new "Dojo" system will train complex neural networks for its autonomous cars. The company's custom D1 accelerator yields industry-leading performance and I/O bandwidth. It implements a unique method to combine 25 chips in one package, delivering more than 9Pflop/s (9,000Tflop/s) of peak BF16 throughput. Each Dl chip has 354 custom CPU cores and employs hundreds of serdes lanes to achieve its industry-leading chip-to-chip bandwidth. To train its vast neural networks, Tesla must connect thousands of these chips. It packages them using TSMC's latest advance in integrated-fanout (InFo) technology, which employs a wafer-size substrate to support and connect more than two dozen chips, increasing compute density and simplifying deployment. TSMC packages and manufactures the design in its 7nm technology and has already delivered engineering samples, which run at up to 2.0GHz. The carmaker designated the product for internal use only.