Recently, Jensen Huang, the CEO of the American chip company NVIDIA, unveiled the AI chip B200 at the 2024 Developer Conference. It boasts a computing speed 30 times faster than its predecessor, earning Huang the moniker of "the most powerful AI chip in the world."
B200 chips have significantly reduced costs and energy consumption compared to the first-generation chips. Previously, training the ChatGPT chatbot for three months required 8000 chips and 15 megawatts of power. Now, it only requires 2000 chips and consumes 4 megawatts of power.
Today, the minimum size of transistors, the basic units of chips, has reached below 4 nanometers (1 nanometer = 0.000001 millimeters), almost approaching the diameter of an atom, reaching a limit.
Developing artificial intelligence requires more powerful chips. We can combine more chips together to form larger virtual chips, constructing massive AI supercomputing clusters. This is the architecture of Nvidia's next-generation chips.
The first step in the new architecture is to combine two chips into one, forming the B200 chip, which totals 20.8 billion transistors, setting a new record in transistor integration. Its memory capacity has also doubled.
The combination of two B200 chips with a standalone Grace chip forms the GB200 superchip, interconnected via ultra-low-power interconnect technology.
Two GB200 super chips assembled onto a single motherboard form the architecture of an artificial intelligence computing node. When 18 such computing nodes are connected, they constitute a larger virtual chip.
The performance of such computing units is on par with the entire supercomputer cluster from the era of the previous-generation NVIDIA H100 chip. This is what many media outlets refer to as "one cabinet tops one cluster," illustrating the significance.
While reports suggest that Nvidia's B200 chips may come with a hefty price tag, tech companies worldwide are lining up to place their orders, with no guarantee of immediate availability for every company.
