This is the last HPCwire issue of 2006. I thought I'd take some time and quickly recap some of the bigger events and trends of the year. There certainly was no shortage of stories.
The Empire Strikes Back
Intel rebounds against AMD with a brand new processor microarchitecture. The Xeons finally achieve performance parity (and then some) against the current crop of Opterons. Intel also brings us 65nm chips and the first (sort of) quad-core processor. AMD swallows its pride and ATI in one gulp and proposes to change the game with its newfound GPU division.
The supercomputer maker has one of its best years in recent memory, gathering several high-profile supercomputing wins and capping the year with the DARPA HPCS selection. “Adaptive Supercomputing” opens to positive reviews.
The Fall and Rise of Silicon Graphics
SGI crashes, burns and resurrects itself in the space of a few months. Even the business cycles are accelerated in HPC. Best of luck in 2007.
InfiniBand gets some new respect from the HPC cluster crowd this year, jumping to 20 Gbps to remind everyone why 10 GbE is not the ultimate answer to HPC interconnects. IB is still a question mark in the enterprise though. We'll see what the OEMs do.
With the help of $10 billion from the Itanium Solution Alliance and the introduction of the dual-core Montecito processor, the Itanium is poised for stardom. At least that's what Intel, HP and IDC are telling us. 2007 may be a make or break year.
DARPA the Decider
After a six month delay, DARPA makes its HPCS Phase III selections, rewarding both IBM and Cray with a quarter of a billion dollars to reinvent supercomputing. To IBM, that's just lunch money, but for Cray, that's serious coin. One question: “What are you guys really building?”
Sun takes its “Proximity Communication” technology and goes home.
HPC Gets Bored with CPUs
Techies decide CPUs can't be the answer to everything. Accelerators come into vogue.
Cell: Mercury Computer Systems and IBM unveil Cell-based HPC gear in 2006. Cell will go into the Roadrunner petaflop super at Los Alamos. Popularity cuts both ways — it's easier to buy a Cell blade than a PlayStation 3.
GPUs: GPUs may be the next general-purpose processing engine. AMD and NVIDIA develop products for the budding GPGPU market. Can Intel be far behind?
FPGAs: HyperTransport enables some interesting solutions. Still working on the programming model though — hope springs eternal.
ORNL and LANL order the first two petaflop-class machines for 2008. The only real surprise: neither one is Blue Gene. In Japan, Riken announces a special-purpose petaflop machine, MDGrape-3.
The 800-Core Gorilla
The HPC community obsesses over multi-core. The consensus: “We're not ready.” That's OK. Neither is the rest of the industry.
Till Next Year
I'd like to thank our readers, contributors, sponsors and the entire staff at Tabor Communications for all their support this year. I'd also like to show my appreciation to all the media relations folks who fill up my Inbox every day with wondrous tales about the most inspired group of people in the IT industry — the HPC community. As for me, I will be taking a much-anticipated two-week holiday break. Our next issue will be published on January 5th, 2007.
As always, comments about HPCwire are welcomed and encouraged. Write to me, Michael Feldman, at [email protected].