Bill Moffitt's Weblog

Wednesday March 16, 2005

The next big fun thing
As promised, I have turned my attention in a completely new direction: hardware.

The amazingly successful rollout of Solaris 10 (over 700,000 registered licenses at the time of this posting), I wanted to take a look at what is the "Next Big Fun Thing" to do.

I considered a lot of different options, both within and without Sun, and I decided that the most exciting thing coming up in the industry in the next year is actually inside of Sun. If you haven't already, I want to encourage you to get familiar with a project inside of Sun called "Niagara." Here's a good external view.

Niagara has been a long time coming; it is the culmination of the work that was being done by Afara, which was acquired by Sun in 2002.  It's a really brilliant idea that was way ahead of its time - these guys foresaw the current situation, where actual performance is not increasing anywhere near linearly with increases in clock speed. So they wanted to add a "third axis" to radically increase the performance of CPUs given a much slower ramp in clock speed.

What they came up with is the processor currently code-named "Niagara," with up to 8 cores per die with each core capable of executing 4 threads. Contrast this with IBM's Power5, which has two cores per die and can execute two thread per core, and the "HyperThreaded" Pentium, with a single core and the ability to sometimes execute two threads simultaneously, and you can see how radical this is.

That's not the end of the radical ideas in Niagara, though. At a time when state-of-the-art processors are drawing well over 100 watts, Niagara is designed to run all eight cores on a bit more than half that. This translates to the opportunity for unparalleled compute density by almost any measurement: MIPs per rack unit, MIPs per Kilowatt, MIPs per ton of air conditioning, etc. With datacenters and service providers being pushed to deliver more and more computing and communications power, and this has the potential to completely change the economics of the computer industry.

Unsurprisingly, Niagara is a set of compromises. To deliver 32 threads, the cores have been simplified with single instruction pipelines, in-order execution, and only one floating-point unit shared across all the cores on the die. The initial clock speed is going to be considerably slower than the "state of the art" processors being sold today. One of the useful visuals we are using is waterfalls: increasing clock speed is like a higher waterfall, while adding threads is like a wider waterfall. There's a lot of power either way, but it's intuitively clear that there's a lot more energy coming out of a wide waterfall than a high waterfall.

Sun's Opteron systems (current and future) are very exciting, but don't set us radically apart from our competitors - everyone has access to Opteron and the new EM-64T Xeon. Sun's UltraSPARC and Fujitsu's SPARC64 systems (current and future) are very interesting as direct competitors to HP and IBM's enterprise-class machines, but they are still relatively conservative - two and four threads, increasing clock speeds, etc. Niagara is completely different, and, properly used, it can deliver an order of magnitude better price-performance to Sun's customers.

My job is to show the world how, where, and why.

Look for more...
(2005-03-16 11:33:25.0) Permalink Comments [1]