By 2030, Russia wants to build ten supercomputers, each with half an exaflop of computing power, using 10,000 to 15,000 Nvidia GPUs.

The Russian project aims to push the boundaries of computing capabilities by building ten new supercomputers, each of which would be among the ten fastest machines in the world today. This should give the nation enough computing power to train various AI models like ChatGPT.

There would be space for 10,000 to 15,000 Nvidia H100 GPUs per system, good for 450 FP64 teraflops, almost half an exaflop. Only the US reaches this level of supercomputers today with Frontier (1,685 FP64 exaflops), the world’s first exascale system already active. Jupiter, the first exascale system in Europe, will be operational next year.

According to Tom’s Hardware, the cost of building ten such supercomputers today would be about $6 billion. By 2030, these costs could potentially be much lower, around $500 million to $700 million.

Smuggle tens of thousands of chips?

It’s great that Russia is pursuing such projects, but since the war with Ukraine there has been a ban on trade with the USA for technological components. How will Russia get around these export rules? Smuggling tens of thousands of very expensive Nvidia GPUs across the border doesn’t seem natural to us. The Russian government does not show how it intends to solve this problem.

Today, Yandex owns the most powerful supercomputer in Russia with 1,592 nodes that have Nvidia A100 GPUs. With a computing power of 21.53 petaflops, the system ranks 27th in the world. Three of the seven Russian supercomputers included in the top 500 list belong to Yandex. Sberbank (two), MSU (one) and MTS (one) control the rest of the systems. Each of these systems has Nvidia chips on board.

Worldwide, Russia is now ranked 12th on the Top 500 list. The USA leads the list with 150 machines, followed by China (134), Germany (36) and Japan (33). The Top 500 list is updated every six months. The most current figures are from June 2023.