Tesla Says It Has Increased The Number Of 80GB Graphics Units In Its Supercomputer To 7,360. It Makes The Tesla Supercomputer Among The Top Seven Supercomputers In Terms Of The Number Of Chips.

According to Tom’s Hardware, Tesla has added “thousands” of Nvidia A100 graphics chips to its supercomputer to increase its power. Until a year ago, Tesla’s supercomputer was equipped with 5,760 A100 chips; But now it has 7,360 of these graphics chips.

It means that the number of graphics chips of the Tesla supercomputer has increased by about 28% in recent years. Tesla’s engineering director says recent changes will make the company’s supercomputerElon Musk is among the top seven supercomputers in the world in terms of the number of GPUs.

Nvidia A100 is one of the powerful graphics chips based on the Amper architecture, which is designed for data centers. This chip uses the same graphics architecture as RTX 3000 series products. Of course, among the consumer graphics of the RTX 3000 series, there is no model at the same level as the A100. The A100 graphics card is equipped with 80 GB of HBM2e memory and offers a bandwidth of up to two terabytes per second.

The A100 is a 400W chip, and its architecture has undergone improvements to optimize the chip for AI processing, data analysis, and high-power computing (HPC).

The first system that used the A100 chip was the Nvidia DGX A100, which used eight chips. When introducing DGX A100, Nvidia said that this system achieves ten petaflops in INT8 format, five petaflops in FP16 form, 2.5 teraflops in TF32 format, and 156 teraflops in FP64 format. The eight chips mentioned above were connected through six NVSwitches and provided a two-way bandwidth of 4.8 terabytes per second.

The numbers we mentioned are for 8 A100 chips, While Tesla’s supercomputer has 7,360 of these chips.

Tesla has not released benchmarks for its supercomputer. The NERSC Perlmutter supercomputer with 6,144 A100 chips achieves 70.87 petaflops in the Linpack benchmark. By examining such data and other information, some media say that Tesla’s supercomputer can probably provide about 100 petaflops of power in Linpack.

Tesla does not intend to continue using Nvidia graphics chips in its next supercomputers. In 2020, Elon Musk said Tesla was working on the Dojo supercomputer and unveiled the D1 Dojo chip last year. This chip is used in Tesla’s supercomputer to increase significantly processing power and throughput.

The Tesla Dojo D1 chip is a type of ASIC specially made for training artificial intelligence and is considered one of the first ASICs in the world of supercomputers. Current D1 prototypes are manufactured with TSMC’s 7nm lithography and have around 50 billion transistors.