x86 continues to dominate HPC

This morning, I saw that a new edition of the Top 500 Supercomputers list was released. This blog entry contains the highlights. Of note to me was the number of system that were using x86:

A total of 354 systems (70.8 percent) now use Intel processors. This is up from six months ago (289 systems, 57.8 percent) and represents the largest share for Intel chips in the TOP500 ever. The AMD Opteron family, which passed the IBM Power processors a year ago, remained the second most common processor family with 78 systems (15.6 percent), down from 105 systems (21 percent) six months ago. 61 systems (12.2 percent) use IBM Power processors, down from 85 systems (17 percent) six months ago.

If you total the Intel and AMD contributions, you'll find that x86 represented more than 86% of the Top 500 supercomputers. If there is any doubt that "commodity" hardware is conquering the world, this should set you back a bit. It's true that the #1 and #2 systems on the list are non-x86 systems (PowerPC-based IBM BlueGene, to be exact), but it's clear that systems based on mainstream processors are the norm for those optimizing the cost/performance tradeoff.

I think the situation in high-performance computing (HPC) is indicative of things to come in networking. The #1 position on the Top 500 list is a very purpose-built system using unusual hardware for the utmost in performance. As you go down the list, most of the systems become standard x86 boxes ganged together using Gigabit Ethernet or Infiniband.

In the networking world, we have a couple of different types of systems. We obviously have very high speed L2/L3 switching fabrics. These systems are "fast but dumb," and are all based on ASICs. They will continue to be based on ASICs, but the ASICs will continue to become more commodity, increasingly coming from the likes of Broadcom and Marvell. The other type of system is something I'll call "complex processing," for lack of a better term. These are systems that are "smarter, but slower." They perform deep packet inspection and generally handle higher-level application processing. While these systems may use dedicated ASIC functionality to accelerate specific subfunctions (e.g. crypto hardware), they will increasingly be dominated by x86 processors for the high-level processing.

Of course, Vyatta is developing the software to run those systems and we believe that software will increasingly be open source.