Last week, at its Capsaicin event in Los Angeles, AMD showcased Project 47, a space-saving supercomputer that can achieve 1 petaFLOPS of single-precision compute performance and handle one quadrillion calculations per second.
Though 1 petaFLOPS isn't a record-breaking benchmark for modern day supercomputers that go upto100 PFlops, it is the footprint of this machine that strives to redefine supercomputing into a more affordable reality for machine learning, graphic virtualization, rendering, and other performance intensive tasks. Project 47 leverages Inventec's 2U parallel computing platform P-47, AMD's powerful EPYC server processors and upcoming Vega-based GPUs, all while sitting in a rack of individual servers.
Simply put, AMD's new supercomputer takes 99.93% less space than IBM's $100 million Roadrunner, the first supercomputer that broke the 1 petaFLOP barrier. The supermachine, which was the fastest in 2007, used 12,960 PowerXCell CPUs, 6,912 dual-core Opteron processors and took up as many as 296 racks covering nearly 6,000 square feet of floor space.
On the contrary, the advanced Project 47 only uses 80 Radeon Instinct MI25 GPUs, 20 AMD EPYC 7601 processors, and 20 Mellanox 100G cards, with 10TB of Samsung memory.
But, this system isn't just about footprint and computing performance. It comes with 30 gigaFLOPS per watt energy efficiency and only requires 33.3 kW to deliver one 1 petaFLOP of single-precision or 2 petaFLOPS of half-precision compute performance. As noted by AMD, this is "25% more efficient than competing supercomputing platforms". Team red also claims Project 47 has "more cores, threads, compute units, IO lanes and memory channels in use at one time than in any other similarly configured system ever before".
As of now, there's no word on pricing of Project 47, but AMD says, the system makes an incredible "performance-per-dollar" prospect and should be available from Inventec and their principal distributor AMAX sometime in Q4 of this year.