Hoping to spark enthusiasm for the second generation of its 64-bit chip family, Intel Corp. disclosed results from a series of tests Wednesday which it said illustrate the performance gains on offer from its forthcoming Itanium 2 processor.
Most of the tests were conducted internally by Intel and won't be corroborated independently until systems go on sale in the coming months. But if the results tally with the official numbers when they are released, they appear to support Intel's long-held claim that Itanium 2 will improve on the performance of its first 64-bit chip by 50 to 100 percent, depending on the type of application.
That came as no surprise to one analyst. Intel has applied improvements to Itanium 2's architecture that it learned about -- and wished it could have included -- when it designed the first version of the chip, said Dean McCarron, president of Mercury Research Inc. in Scottsdale, Arizona. "With a second implementation you almost always see a significant performance gain," he said.
Boosting Itanium's performance is essential for Intel as it tries to steal business from makers of RISC chips like Sun Microsystems Inc. which dominate the midrange server market. The first Itanium chip suffered delays before it shipped in May last year and came to be regarded by many analysts as a proof of concept vehicle. Intel hopes that Itanium 2, formerly known as McKinley, will have enough muscle to reach a much wider audience.
"With the first generation of Itanium we did a good job of establishing the architecture," said Jason Waxman, Intel marketing manager for enterprise platforms. "We got the platform out there and got a lot of people to do some initial deployment, some initial evaluation of it. With Itanium 2 the bar is a lot higher for us and the goal is to go out there and establish world-class performance."
Itanium 2 is on track to ship around mid-year along with the first two- and four-way servers and workstations from major Intel customers, he said. Larger systems packing eight or sixteen processors will start to appear gradually in the months that follow, Waxman said. Besides Sun, Intel will compete against IBM and Advanced Micro Devices which has said it will launch its first 64-bit chip later this year.
Benchmarks are about more than bragging rights. Ideally they help customers gauge what type of performance they can expect from a system running a particular type of application. They often include a price/performance ratio, although Intel hasn't disclosed pricing for Itanium 2 yet and so didn't include that with its unofficial test results. The results were due to be presented Wednesday at its Intel Developer Forum in Europe.
Sun was skeptical of the results, which in many cases also claimed to show better performance than servers based on Sun's own UltraSPARC chips. It complained that Intel's results were skewed by the Itanium 2's large, on-chip level 3 cache, which stores frequently used data close to the processor and may have led to performance gains that wouldn't be realized in a real-world setting, according to Sue Kunz, director of marketing for Sun's processor products group. She said the results also overlook that Itanium 2 chews up more power than Sun's chips, which, she argued, makes it harder for system makers to build servers around the chip.
Sun has some cause to be skeptical. Intel was caught out at its developer conference here earlier this year when it contrasted the performance of an Itanium server with what it characterized as a comparable system from Sun. It transpired that the Sun system was based on an older UltraSparc II chip that was about to be retired. "We have to take whatever they tell us at face value," McCarron said of Intel's results. The integrated cache on Itanium 2 might well have affected some of the results, but not all of them, he added.
Intel's tests were intended to measure performance running large databases, ERP (enterprise resource planning) systems and high-performance technical and scientific applications such as those used to design aircraft or model weather patterns. Other tests measured online transaction processing and system memory bandwidth.
For example, Intel crafted what it called an equivalent to the popular TCP-C benchmark to measure online transaction performance. Running databases from Oracle Corp. and Microsoft Corp., Intel said a four-way Itanium 2 server processed transactions 70 percent faster than a comparable Itanium-based system. It had to devise an equivalent test because TPC-C results can't be referenced until they are official, according to Waxman.
To gauge memory bandwidth, Intel said a single-processor Itanium 2-based system delivered 3.7Gbps (bits per second) on the Stream benchmark, which it said is more than 2.5 times that of a comparable Itanium-based system.
For technical applications it ran the Linpack 10K benchmark, also on a four-way system, and said it achieved performance of more than 13 gigaflops, or 13 billion floating-point instructions per second. It claimed that figure almost doubled the performance of a comparable UltraSPARC III system from Sun.
In each case software was rewritten to take advantage of architectural enhancements in Itanium 2. Applications written for the first Itanium but not recompiled for the newer chip would still produce 80 percent to 90 percent of the performance gains, according to Waxman. Systems tested were prototypes built by Intel with 1GHz Itanium 2 chips, which is the clock speed at which the chip will be introduced. The current Itanium tops out at 800MHz.
The performance gains come from a variety of design improvements, Waxman said. Among them, Itanium 2 has a 400MHz system bus that is 128-bits wide, compared to Itanium's 133MHz bus which is 64-bits wide. Itanium 2 has a 3M-byte level three cache that is integrated alongside the processor. The first Itanium comes with up to a 4M-byte cache, but separate from the main processor, which tends to slow data transfer speeds.
Intel expects to follow Itanium 2 next year with two new members, code named Madison and Deerfield, which will be manufactured using a more advanced, 0.13-micron manufacturing process, Waxman confirmed. Further on the horizon comes a chip code named Montecino, which is due in about 2004, he said.