Tesla's In-house AI Processor
The author is an analyst of NH Investment & Securities. He can be reached at hwdoh@nhqv.com. -- Ed.
For Tesla’s D1 processor, packaging technology will be the key process. We recommend paying attention to domestic back-end process players. EDA companies, such as Synopsys and Cadence, should also benefit.
Tesla announces in-house AI processor
During its AI Day event, Tesla announced the Dojo, a self-developed computer for autonomous driving computation, and the D1 processor. The D1 processor, which is manufactured on a 7nm process, is 645mm2 and integrates 50bn transistors. It supports new formats, including FP32, BFP16, and CFP8, with BFP16/CFP8 computational performance standing at 362 TFlops and FP32 performance coming in at 22.6 TFlops.
The D1 has unique packaging technology as multiple processors are connected without bonding. In detail, 25 D1s are connected with an interface and power & thermal management modules to form a Training Tile. Individual processors are mounted on silicon interposers that support high-bandwidth interconnects and are equipped with heat plates with coolant flowing on top. The Dojo system consists of 3,000 D1 processors and 120 tiles.
D1 processor introduces new concept in packaging technology
There are already numerous companies that have developed their own machine learning CPUs, including Google, Amazon, and Baidu. Thus far, Tesla has used a general-purpose processor, Nvidia’s A100, for its autonomous driving analysis. However, the D1 should be greater optimized for autonomous driving calculations than a general-purpose processor.
Compared to existing ASICs, the D1 introduces a new concept in packaging technology rather than front-end processes or design processes, such as interprocessor connection and expansion or thermal management. This is in line with the current trend of solving the limitations of front-end process technology with back-end processes. We advise paying attention to domestic back-end process players such as Nepes, SFA Semiconductor, and LB Semicon. EDA companies, including Synopsys and Cadence, should also benefit from the emergence of new in-house developed processors.