From MIPS to exaflops in mere decades: Compute power is exploding, and it will transform AI
1 min read
Summary
Nvidia unveiled a single rack system of servers capable of one exaflop operations per second at its GTC conference, marking what the company calls a “breakthrough” in computing density, energy efficiency and design.
While the Nvidia system has achieved roughly 73 times greater performance density than the Frontier system in three years, it is geared towards speed over precision using AI workloads, using 4- and 8-bit floating point operations, while Frontier used 64-bit for scientific simulations.
The development has led to questions about the future need and benefit of such computing power, and whether AI dta centres are being overly ambitious with Project Stargate alone envisioning 20 data centres across the US, with others in the planning stage.
However, computing power is likely to surge with the advent of AI agents and digital twins, and $40 billion of new funding for Open AI promises “ever more FLOPS coming online”.