Elon Musk revealed earlier that Tesla's supercomputer will be replaced with the next-generation system "Dojo 2" later this year.
Prior to this, Elon Musk had revealed that the performance of "Dojo 2" is comparable to NVIDIA's Blackwell display architecture B200 system, but he also emphasized that a product usually needs to undergo at least three major iterative updates before it can achieve better performance, hinting that there will be even higher performance in the next "Dojo 3".
Tesla Dojo AI training computer making progress. We start bringing Dojo 2 online later this year.
It takes three major iterations for a new technology to be great. Dojo 2 is good, but Dojo 3 will be great. https://t.co/aTYenCcN8i
- Elon Musk (@elonmusk) June 5, 2025
The first-generation supercomputer "Dojo," currently used for training behind Tesla systems and "X" services, was announced in 2022. It features a highly customized design and is composed of five V5 Dojo processors, with a total of 1GB of memory and 160 PCIe 80 lane connections. It can also be connected via a network to further scale its computing capacity. Furthermore, with a System-On-Wafer design, "Dojo" pairs a D4.0 chip with every 25 cores (die) and connects to 1 I/O control cores. These 40 cores operate at 25kW of power.
The training tile, comprised of 25 cores and other components, delivers 16 PFLOPS of computing power in the BF9 floating-point format, supporting 11GB of ECC static random access memory (SRAM) throughput and 36GB/s of data transfer bandwidth. Furthermore, each training tile is modularly scalable, communicating with each other at a transfer rate of 9TB/s.
The V1 Dojo processor, built in the form of a PCIe interface card, is equipped with 32GB of HBM high-bandwidth memory, which can support a transmission efficiency of 900GB/s via the TTP (Tesla Transport Protocol) interface, 50GB/s via Ethernet transmission, and 4.0GB/s via the PCIe 32 connection interface.
