Recent years have seen an increased interest and applications in artificial intelligence. The Dojo D1 chip is Tesla’s newest ASIC for AI training, designed to accelerate the company’s activities in this area.
The number of companies building ASICs for artificial intelligence computing is increasing, from startups to giants such as Amazon, Biesta, Intel, Nvidia, etc. It has led Tesla to develop its own ASIC for training artificial intelligence because not all of them work properly.
The chip, called D1, is used by the Dojo supercomputer, which will be used to train a lot of artificial intelligence models in the future. The Tesla D1 has 50 billion transistors and Tesla uses TSMC’s 7nm lithography. The dimensions of this chip are 645 mm2.
In order for Tesla to perform well, we have to expect a high level of efficiency. This results in an accuracy of FP16/CFP8 of 362 teraflops based on the D1 output. Tesla’s chip performs better than Nvidia’s most powerful GPU, the A100, for FP16-type data. With its product, Nvidia offers a traflapse of 312 for FP16.
Tesla has created a silicon chip based on a network of functional units (FU). FUs contain a dedicated ISA and 64-bit processor designed to transmit, collect, and play navigation information.
Each FU can have a power of 1 Traflapse in BF16 or CFP8 calculations and 64 GAflaps for FP32 calculations. Additionally, it has 512 GB/s of bandwidth in both directions. Generally, there is a decreased amount of delay and increased performance due to its design.
