---

At that time, [ten years ago] the world’s fastest supercomputer in the world was Sandia’s ASCI Red (ASCI on Top500.org, picture) with 1.338 Teraflops sustained throughput. It was a massive structure with a huge array of servers consuming 2500 square feet, 104 cabinets, 76 computers, 8 networks and 20 disk units. It had 2 PentiumPro CPUs per node, with 4510 compute nodes. It was a monster!

Today, AMD’s GPU today has surpassed its performance with 1.36 Teraflops of compute on 190 watts max power. Phenomenal. [By the way, it's worth noting that in 2007, I heard one of the leading multi-node architects at Intel give a lecture where he discussed that even when compared with modern CPUs (of 2007) on extremely memory intensive applications, ASCI Red was still a contender in throughput due to its custom developed network fabric and high memory bandwidth. Even though it was powered by meager 200 MHz Pentium Pro CPUs (originally), which are but a drop in the bucket compute-wise next to today's high-end Xeons, it's the interconnect that made the machine so sustainable across huge, massively parallel, bandwidth hungry apps, even when compared to modern CPU-based supercomputers.]

---