From the San Jose Mercury News:
Posted on Sun, Apr. 27, 2003
Thread designs divide chip makers; should they be skinny or fat?
By Dean Takahashi
Mercury News
The makers of the world's fastest microprocessors are engaged in a theological debate: Which are better, fat threads or skinny threads?
The high priests at Sun Microsystems, Intel, IBM and Advanced Micro Devices are betting on designs that will determine who will control the multibillion-dollar server computing market.
All agree that a technique called chip multithreading, or getting a chip to execute several independent programs (threads) simultaneously, will be useful in an era where they have hundreds of millions of transistors at their disposal. But they disagree about when and how.
Sun is betting on skinny threads, which break a program into many parts. Each part is executed more slowly, but at the same time. By pouring resources into skinny threads, Sun thinks it can outrun those focusing on what amount to fat-thread dragsters.
``The others were racing to build the fastest processor, but they were running the wrong race,'' said David Yen, executive vice president for Sun's Sparc chip division. ``The others cannot so easily abandon the paths they were on. By going our way, we think we can simplify computing and get a lot more work done.''
IBM, Intel and AMD have put their money on fat threads, or processing one or a few software programs extremely fast. They recognize the value of adding more threads, but they say Sun's approach comes at the expense of performance on a single thread.
``Sun is doing something extreme because they are struggling to find something that no one else is offering,'' said Ravi Arimilli, an IBM fellow and chief server architect. ``They are losing their high-end market share and looking for new territory.''
Yen's architects believe that memory latency, or the delay that results when a fast processor has to stop what it's doing to fetch data from a slow memory chip, is the main bottleneck in today's servers. Server processors have become like a Ferrari engine that runs at 150 mph but can't get enough gas to keep running.
To tackle this problem, Sun is designing chips that, when they debut in 2005, will have four processors that each handle four threads, or the subprograms that can be executed independently. This way, whenever a processor stalls because it has to fetch data, it moves from one thread to another.
``Memory has been falling behind for a long time,'' said Yen. ``Rather than fight it, we're learning to deal with it. It's a fundamental change of mind.''
Reversing the trend
Each of Sun's processors won't be as speedy as Intel's Itanium or IBM's Power chips. But Sun has long given up on dominating computing on the desktop.
The bet is a crucial one for Sun, whose UltraSparc chips have lost ground to Intel's 32-bit Xeon and 64-bit Itanium chips. Observers say that Sun, which has reported eight quarters of weak financial results, needs something to reverse the trend.
``If this gives Sun's customers hope, the customers will hang in there,'' said Nathan Brookwood, an analyst at Insight 64 in Saratoga. ``I believe Sun's theory makes a lot of sense and they are better positioned to pursue this path than a lot of folks.''
Sun has always had to justify why it spends hundreds of millions a year on chip research when big rivals like Hewlett-Packard are relying on Intel for next-generation server microprocessors. The pressure has been rising as customers defect to Intel.
``2005 is a long way away,'' said Gary Campbell, vice president of strategic architectures at Hewlett-Packard, which has teamed up with Intel. ``They have a weakness before then.''
Strategy
The way that Sun came to its new religion on chip design doesn't necessarily inspire confidence. Marc Tremblay, a chip architect, proposed a new architecture called MAJC that used multithreading in 1999. That chip was geared for multimedia and was aimed at outdoing the graphics performance of Silicon Graphics. But it didn't take off and Sun didn't immediately apply the lessons to designing server processors. Meanwhile, Intel launched Itanium in 2000.
``If Sun had taken the people working on MAJC and put them on a server processor, they might be ahead of the game now,'' said Peter Glaskowsky, editor in chief of the Microprocessor Report. ``I don't think it's too late now, but they could have done this better.''
Instead, it took a start-up to force Sun to solve its internal debate about how to move forward. Les Kohn, a former Sun chip architect and Stanford engineering professor Oyekunle Olukotun started Afara Websystems in 2000 to create multithreaded Sparc microprocessors. While other chips had become hopelessly complex in their quest for speed, Afara felt that simpler processors running lots of threads would be faster.
In July 2002, Sun agreed to buy Afara for $28 million. That ended the debate inside Sun.
The company plans to launch a 16-thread chip in 2005 and a 32-thread chip in 2006. Such chips fit with Sun's software strategy, since its Java programming language and Solaris operating systems are far better at handling threads than rival technologies. Linux and Windows, for instance, are poor at multithreading.
Yen says that Sun is now ahead of the game because its rivals' chips are much larger and more complex. It's harder to put lots of such processors on one chip, and harder still to make them run multiple threads. By 2005, Intel may only be running two or four threads, based on its current plans.
All at once
Kohn argues that server tasks often depend more on how much work can be done simultaneously. An example is a computer that serves Web pages to Internet users. Each time a customer logs on to a Web site, a thread is sent to a processor to send Web page data. Those customer requests can be processed in parallel much faster if a computer system processes lots of threads at once.
To some degree, multithreading is becoming more and more fashionable. IBM is using it in its new cell microprocessors for Sony's PlayStation 3. Most new network processors, including one from Intel, are using it for communications processing tasks. Advanced Micro Devices says it will use chip multithreading as it becomes practical in its server processors.
``Multithreading is the next big wave,'' said David Fotland, a former Hewlett-Packard microprocesor architect and now chief technology officer at Ubicom, a chip maker that is using multithreading in its microprocessor for wireless networking tasks. ``It's going to happen everywhere.''
If threads really are so good for server applications, Intel and IBM say they will be able to respond. Intel already has two threads running on its Pentium 4 chips and it has 128 threads running on its network processors. But those network processors are running specialized software, and it will take some time to switch gears for servers.
``We believe we don't need to pull the trigger sooner because we are already leading in industry performance,'' said Nimish Modi, general manager of Intel's enterprise processor division.
Kohn believes that other factors, like better power consumption, will force the adoption of multithreading. He also notes that stamping out small processors on a chip will simplify design, which is becoming risky with growing complexity. Sun can design just one processor and replicate it automatically across the chip.
Sounds simple, but Sun needs to catch up now. Intel is launching its third Itanium chip this summer and IBM is launching its Power 5 chips next year. Meanwhile, Sun is launching UltraSparc IV, which has two processors on one chip, later this year. As an insurance policy, Sun is extending the UltraSparc line in case customers want meaty fat-thread processors.
``It's hard to get a jump on Intel,'' said Linley Gwennap, an analyst at the Linley Group in Mountain View. ``Intel can spend money to catch up. Since Intel has the technology today, it's hard to see how far ahead Sun is.''
--------------------------------------------------------------------------------
Contact Dean Takahashi at dtakahashi@mercurynews .com or (408) 920-5739. |
