Reality Check | Ephraim Schwartz » Intel has a gripe with the software industry. Is Moore's law becoming a software afterthought?

May 25, 2007 | Comments: (0)

Intel has a gripe with the software industry. Is Moore's law becoming a software afterthought?

TAGS: None

The saying "the more things change, the more things stay the same" was never more evident on Friday when Intel and Sun sponsored a roundtable discussion on hardware and software development for the press in San Francisco .

The change was represented by the fact that Intel no longer makes single-core processors in favor of only multi-core processors. What has stayed the same, at least for the past 12 years that I've been dealing with Intel, is the Intel mantra: The software has to catch up to the performance of the hardware.

Back then, Intel used a subtle approach. Whenever the chip maker unveiled a new chip, it would have its software partners on hand to demonstrate applications of the future that would use every cycle of processing power put out by the latest processor.

At the roundtable discussion, however, Intel was more into the blame game. This time Shekhar Borkar, Intel Fellow, director of microprocessor research, expressed his obvious frustration by saying he had a "complaint" with the software industry for not creating applications that can take advantage of parallel processing.

"It is imperative that software has to double the amount of parallelism [it can address] every two years," Borkar said.

However, when asked later in the discussion what will happen if they don't, his answer was weak. He said if one company doesn't do it, a competitor will.

That may be true but two things seemed even more obvious to me.

One, the software industry has a lot more on its plate than parallel-processing-enabling its applications.

The industry is going through great change at the moment, mainly with regard to Web 2.0, and here the network and performance through the network is driving much of software development. See companies like Akamai, and Citrix who are looking at ways to deliver and execute applications in the network, and Adobe, whose Apollo project is looking at a way to create hybrid applications that start in the network and finish on the desktop.

The other obvious truth is that this is not a technology problem for the software industry but a business problem.

Intel's business model requires that it must sell its chips in the millions and so far the only way they have come up with to do that is to tout the chips' performance, and to a lesser degree, the power savings on mobile devices.

I contend that if the software industry had the same business model, it would have solved the problem of how to use parallelism in its applications years ago when it was first briefed by Intel and others on this technology.

Rather, the software industry has a different business model, and other problems to solve that include inter-operability, integration, SOA, ease-of-use, consumerization of the enterprise, Web 2.0, software as a service, master data management, compliance, e-discovery, the list goes on. While performance is always needed and appreciated, the goals are far more complex.

I think it is disingenuous on the part of Intel to chastise the software industry as it did at this roundtable when it is obvious Intel has a vested interest in making performance the goal of every segment of the high-tech industry.

Nevertheless, Borkar continued to chide, saying that since Intel comes out with a new multiple of its multi-core processor every two years, the software industry has two years to go from 2-way to 4-way, and two years again to go from 4-way to 8-way.

In a moment of candor, Borkar did say, "If the software doesn't keep up, the chip becomes a paperweight."

One of the more interesting solutions that may not require software to keep up with the pace of hardware performance also came from Borkar, however.

It is called domain specific languages.

For example, you could write a program using a domain specific language just to address TCP/IP processing. And, if you want to write a program to parallelize TCP/IP you could write it just for that kind of network.

For further enlightenment I turned to one of the best and most widely quoted chip analysts, Nathan Brookwood of Insight 64 to gain a better understanding of the problem of creating multi-threaded aware applications.

Brookwood complied, in spades.

"To think how to code in parallel threads is an unnatural act," Brookwood told me.

By and large we think in a linear fashion, Brookwood said. Clearly grammars to think and express algorithms in parallel form are needed but most tools don’t do that.

Brookwood believes we are going to have to wait for the current generation of linear programmers to die off or get kicked upstairs into management and before we get a new generation that thinks parallel.

There is, however, a Sun tool or parallel development language, not quite productized yet, called Fortress that allows programmers to design applications that can take advantage of multi-core processors.

"When you write stuff [in Fortress], it is inherently parallel and if you want things not parallel you would have to go out of your way," Brookwood said.

The problem really is that multi-threading only helps a few computer applications that appeal to ordinary people. Sure it is needed for scientific computing but remember, as I said, Intel needs to create a demand among millions of buyers, not thousands.

The killer app does not yet exist and even Borkar admitted that.

Nevertheless, Brookwood does believe that over time it is important to create programs that are multi-core aware.

Up until now software developers had a free ride. Intel and AMD could make one, single, faster chip and the software ran faster. That kind of free ride ran out of steam and the new ride is multi-core.

"Multi-core is the new megahertz," Brookwood said.

However, it is more visible to the software than megahertz. It requires the programmers to think about and change what they are doing.

Unfortunately, that really hasn't happened. As Intel went from a single to dual thread, the ISVs created a whole bunch of narrow solutions that work with two threads, but they didn't think in terms of the bigger picture: multi-threading as a concept.

Now these same developers have to rethink how to make an application work from two threads to with four threads.

In the meantime, Intel doesn't want to wait five or ten years to make sure its processors don't turn into paperweights.

But I think they may have to wait, and over time the industry may take us in so many new directions Moore's law may become a mere afterthought.

Posted by Ephraim Schwartz on May 25, 2007 01:59 PM

RATE THIS ARTICLE: