Sun: We Can Build a Faster Supercomputer Than IBM

Sun Microsystems announced this morning it has developed an architecture for supercomputing which, as soon as the fourth quarter of this year, will drive a data cluster capable of a sustained processing rate of 1.7 petaflops. That's six times faster than the machine perched atop the most recent Top 500 supercomputers list today...and IBM has already reacted.

IBM doesn't like to be beaten in the race to a milestone, by Sun, Intel, or anyone else. When Intel announced it had discovered the long-sought-after formula for high-k-plus-metal-gate semiconductors, IBM made pretty much the same announcement just a few hours later, claiming it had actually reached that goal a few hours before.

Now, IBM has made claims circulating among blogs this afternoon that its new BlueGene/P supercomputer has already bested its own existing champion, BlueGene/L, in sustained operations per second by a factor of four.

While the University of Mannheim - which publishes the Top 500 list twice a year - may yet verify IBM's claim, its timing is an indication that something's up. Sun Microsystems believes it may yet beat IBM to the one petaflop milestone - one thousand trillion instructions per second, sustained speed - with a device it's building for the Texas Advanced Computing Center.

Using an architecture Sun calls Constellation, one critical element of the new system is its switching capability between processor cores. Using a radically redefined InfiniBand system, Sun aims to reduce the number of switches in a supercomputer data cluster from the typical 300 - all of which would be continually communicating with one another - to one.

And while Sun is talking about UltraSPARC processors as likely providing the power for the first TACC Constellation model, the company insists on maintaining open architecture whenever and wherever possible. To that end, it will be offering TACC and other customers the ability to install UltraSPARC, AMD Opteron, or Intel Xeon blades.

Right now, the options list mentions as much as an eight-core UltraSPARC T1 or a quad-core Xeon, but only a dual-core Opteron, though AMD may yet produce a quad-core that fits the Constellation architecture.

The goal post for Sun is set at 1.7 petaflops. And as Sun's blade server product line director Mike McNerney told reporters yesterday, "The TACC solution, when it comes online, is expected to be the world's largest supercomputer."

That goal may be even harder to reach if one of Sun's statistics is borne out by independent testing: Memory latency is recognized to be the greatest obstacle among system designers preventing throughput gains even as more processor cores are added to a supercomputer's data processing cluster. Some have said it's like accelerating an electron further and further within a cyclotron: There comes a point where you just can't get it any faster no matter how hard you push.

Sun's 700 ns memory latency claim is maybe a bit too transparent. That would be a huge hurdle for the company's architects to overcome. But it could be the InfiniBand switch that makes it possible, overcoming latency issues by reducing inter-processor traffic by orders of magnitude.

If this were the automotive industry, the supercomputer architecture part of the business would be Formula 1 development. It's the part where manufacturers claim they're making the technologies that will eventually trickle down to the consumer level, though few consumers actually believe this even when that's precisely what does happen.

Sun's contention is that the Constellation is a preview of what's to come: a supercomputer that in a few years' time will become an ordinary enterprise data cluster. And the window of time between supreme and ordinary is ever narrowing.

"We're starting to see this idea of, how long does it take from conceiving, funding, starting to build a supercomputer, and when it actually turns on and people are able to develop and run their applications against it?" asked McNerney rhetorically, representing consumers whom he claims are driving this transition.

But by how much? If the typical industry benchmark transition window is five years, are we talking about shaving several months off of that, perhaps a few years?

"I actually believe it's quite a bit narrower than that," McNerney responded to BetaNews. "When you start to look at the enterprises out there, especially a lot of the customers that we deal with that are really leveraging IT as a competitive advantage, they're starting to see a lot of these same issues, maybe not at the scale you see in the HPC environment, but I would say they see them in the same timeframes."

It took McNerney a few minutes, but eventually he got around to some numbers: "The complexity of building out big clusters, if you go down and talk to eBay or financial service customers, they certainly are facing those same issues today. I don't think they face them to the same degree, so their need to solve them is perhaps not as high, but I actually think those customers would argue with you about that. They would love to see these kinds of simplifications, so I actually think that timeframe is compressing quite a bit, and I would say it's more in the 12-month timeframe at the outside, versus five years."

That's a one-year window between super-performance and ordinary performance. Does that mean a similar window awaits customers of enterprise-class data clusters, between high-performance and obsolescence?

We didn't expect an immediate answer to that, so we found a different way to approach the question: First, we changed the subject to talk about the cooling problem. McNerney spoke about Constellation as a framework architecture for innovative cooling in blade systems across the board from Sun. With thermal design point ratings of 60 W for UltraSPARC options to as high as 120 W for some AMD Opterons, how does Sun plan to maintain optimum temperatures with such variations in core temperature?

"I think power and cooling is a very confusing topic out in the marketplace, and trying to do apples-to-apples comparisons on power and cooling becomes quite difficult," McNerney responded. Being able to implement a blade architecture in and of itself, he said, gives builders more room to deploy larger power supplies and fans.

Of course, the real problem occurs when manufacturers such as AMD and Intel lower their thermal design points, which typically happens within 12- to 18-month timeframes. Here is where timeframe re-enters the discussion: Customers want to be able to take advantage of that cooler processing, especially in high-performance environments. So when do they start to feel the window closing?

"When we look at that, we actually believe we have potentially a five-year window on the power and cooling capacities that we see in the roadmaps from the major vendors," McNerney told BetaNews. "So we'll be able to take this blade chassis and upgrade it to the latest and greatest technologies for the next five years, without having to go back and update and change your data center - which, when you think about deploying this massive data center with all these compute nodes, is a significant undertaking."

So maybe one year's time for shelf life toward market obsolescence, but a five year window of useful deployment life in the laboratory. That actually isn't a change from the industry standard.

"Texas Advanced Computing Center might be the largest supercomputer on the planet," he continued. "This isn't just targeted at them. This is actually targeted at bringing this level of cluster computing into a more mainstream environment, if you will, your own personal supercomputer, and being able to run that in your data centers, and product design shops."

People who aren't race fans aren't convinced that making cars go 280 mph on a road course affects them in any significant fashion, other than appeasing their curiosity. But the efficiencies that engineers learn there on the racetrack do get implemented in consumer and commercial automobiles and trucks; the racetrack is the auto industry's space program. Likewise, the supercomputer side of the business is Sun's and IBM's NASA. It's where design innovations get started, and like auto racing, it takes sponsors to make it possible - except here, those sponsors are actually customers, like TACC.

Sometimes the best innovations come when dealing with known restrictions...which is why the 700 ns memory latency Constellation will be dealing with will be an interesting story to follow.

"It's the exciting part of the industry, right?" McNerney told us. "Certainly when you look at this, you'll say, 'Who in the world is going to need 1.7 petaflops of computing?' And the second you ask that, you'll have ten people raise their hand and say, 'When can you do 4 petaflops?' Certainly the demand is insatiable. I think the computer industry has shown that again and again, more performance, more efficiency is critical."