IBM processor with 100 billion transistors for six times faster artificial intelligence

IBM has caused a sensation in the semiconductor industry with the announcement of a new technology. Its experts have managed to design transistors that are just 0.7 nanometers wide. In practice, this means that chip manufacturers could "squeeze" as many as 100 billion transistors onto a surface the size of a human fingernail. This is the first time in the history of computing that someone has managed to "pack" such a huge amount of components into such a small space.

The new 0.7-nanometer chips represent an incredible leap forward. They are 50 percent more powerful and 70 percent more energy efficient than IBM's own 2-nanometer chips from 2021, which were previously considered the smallest in the world.

This shift to ultra-small dimensions will have a huge impact on the field of artificial intelligence. Today's most popular AI accelerators can process around 1,500 TOPS (billions of operations per second). IBM researchers estimate that the new chips could achieve six times that performance, around 9,000 TOPS. In practice, this means that the time it takes to train huge language models, which currently takes around three months, could be reduced to just a few weeks.

The key to this success lies in a new architecture called nanostack. Instead of shrinking transistors on a flat surface, engineers have focused on height, comparing the architecture to building a 100-story skyscraper, while competitors like Samsung and Intel are reaching around 30 to 50 stories in their 3D work.

The team had to overcome several technical hurdles to achieve this feat, including a new way to connect two wafers with minimal defects and increasing SRAM memory by 40 percent, the biggest leap in the last decade and directly addressing the data access bottleneck.

It is worth noting that this achievement is still in the experimental phase, and 3D chip designers still face serious thermal management challenges. Because the layers are extremely thin, heat builds up quickly, which can prevent transistors from turning off properly and ultimately cause the entire chip to fail. Industrial production is still several years away. Widespread use of 2nm devices is not expected until the end of the decade, followed by 1.4nm and 1nm versions, while 0.7nm chips are pushed even further into the future.