Intel to Remove Xeon's Advantages to Push Itanium
by Timothy Prickett Morgan
After nearly a decade of 64-bit processing in the RISC server market, it might be reasonable to accept that 64-bit computing for Intel X86 processors was a foregone conclusion. Since 1996, Intel has been making the case publicly for its 64-bit Itanium and its EPIC (explicitly parallel instruction computing) instruction set. It has not been an easy run for Intel, but the company has a plan to make Itanium take off: Remove the advantages that 32-bit and 64-bit Xeons have on Itaniums, and stress the advantages the Itanium architecture brings.
Whenever a discussion of Xeon versus Itanium began in the past three years, it inevitably ends with a discussion of a 64-bit variant of the Pentium 4 core. The advent of the Xeon-64s, which will begin shipping next quarter with the "Nocona" Pentium 4 Xeon DPs, has not changed the debates one bit--or even 32-bits, for that matter. So let's get this out of the way right now. Itanium is a radically new core that Intel and HP designed for the long haul, and they expected a very long ramp up.
It has been no secret that some of the server vendors--especially those who are not enthusiastic about the jump from the P4 to Itanium instruction sets and/or who have their own RISC/Unix markets to protect--had been rooting for the "Yamhill" Xeon-64s to come to market, and they seems to have used the 64-bit Opterons from AMD to push Intel into getting the Xeon-64s out the door. Many of these same vendors would love to see Itanium go the way of all flesh because Itanium is a real threat to their RISC/Unix bases over the long haul. All of that Itanium bashing has a purpose.
Tom Garrison, who is marketing director for the Enterprise Platforms Group in Europe, and Mike Fister, the group's general manager, dodged the rumors and implications of the Xeon-64s as they discussed the longer-term roadmaps for the Xeons and Itaniums just prior to the launch of the Xeon-64s in February. Intel was then and still is now sticking to its Itanium guns.
The circumstantial evidence for this belief is all over the presentations Intel has made and continues to make. First, Intel is committed to getting the IA-32 Execution Layer for the Itanium chips into production in the first half of 2004. This environment, which is a dynamic translator that will come with future operating systems that converts 32-bit Pentium instructions into 64-bit Itanium instructions. Up until now, the Itanium chip has relied on circuitry in its chips to run IA-32 instructions in a much-degraded mode. With the IA-32 EL, a 64-bit 1.5 GHz "Madison" Itanium can deliver about the same performance as a 32-bit 1.5GHz "Gallatin" Xeon MP. On many workloads, a Xeon running at 3 GHz would deliver about the same performance as a 1.5 GHz Madison, so the dynamic translator is not nearly as good as going native with Itanium or sticking with Xeon. Over time, Intel says that the IA-32 EL should deliver anywhere from 50 percent to 70 percent of the native performance of the Itanium chip when running 32-bit code. The Itanium processors and Intel's own enterprise compilers will be enhanced to make 32-bit code run better.
The IA-32 EL can be downloaded from Microsoft for its 64-bit Windows 2003 operating systems now, which is important since a lot of existing 32-bit code that customers worry about has been created for Windows. When Windows Server 2003 Service Pack 1 comes out in the spring of this year, the dynamic translator will be part of the patch. Linux is the next environment to get the dynamic translator, with Red Hat saying it will get it into its Linux distributions in 2004; SuSE's Linux can't be far behind. The translator was designed to be operating system independent, and only works at the application level, so it should be a snap to make it work on Linux and any other operating system that might be ported to Itanium (like SCO Unix or Solaris X86, should that ever come to pass).
The dynamic translator doesn't so much eliminate one of the advantages of Xeons as concede that its ability to run native 32-bit code is a big issue. If Intel can get the performance of Itaniums ramped up, get the translator running at 70 percent, and cut the costs of the Itanium chips, then it can bring the price/performance of the two processors close to parity on even 32-bit workloads. This is the real issue for customers. People will talk about Xeon-64s until this parity shuts them up.
Intel says that there are other things it will be doing to level the playing field between Xeon and Itanium. Garrison said that the Itanium core is half the size of the Xeon core, which means that Intel will be able to cram twice as many Itanium cores on a single chip for a given chip making process. (This would seem to imply that the Itaniums should be able to run at twice the clock speed or higher, too, but they in fact run at half the clock speed. There are clearly some tradeoffs in moving to the EPIC architecture.) Still, in 2004, Intel expects that the Itanium will show a 30 to 50 percent performance advantage compared to Xeons on a per processor (not per core) basis for enterprise (database and applications) and technical workloads. And by 2007, the performance advantage for Itanium compared to Xeons (again, per chip, not per core) will be on the order of 50 to 100 percent. These comparisons include the new Xeon-64 extensions, which give these chips 64-bit memory addressing. That 64-bit memory only helps performance a little bit, and only for large databases and memory-hungry applications. Faster memory pipes are just as important as wider ones when it comes to servers, as Opterons clearly demonstrate running in 32-bit mode.
One of the last barriers to Itanium adoption is server-level pricing. Garrison says that right now, Itanium machines offer about 30 percent performance advantage compared to Xeon machines for full configurations, but the Itanium machines carry a 30 to 60 percent price premium that still makes a transaction 10 percent more expensive on Itanium than Xeon. Some of that is due to chip costs, that premium is also caused by server chipset costs. Over the next three years, Intel is going to create a single chipset that supports both Xeon and Itanium processors. According to Garrison, this will be the exact same chipset, not a sort-of common chipset like IBM's "Summit" family of chipsets, which have about 85 percent commonality between the Xeon and Itanium versions but which are not identical.
Intel says flat out that the goal for Itanium is for the family of servers based on it to use as many of the same components that there is no disparity in pricing. By 2007, Intel wants Itanium machines to have pricing parity, and to have twice the cores on a die, twice the performance, and twice the bang for the buck.
All of this seems to imply that the days of the Xeon are numbered. And if Intel could have its way, Xeon would no doubt be dead by now. But this transition will take many, many years. How many, Intel is not saying, but it is clearly trying to limit them without undermining its vast X86 business. So when will Intel kill Xeon? "When customers tell us they don't need it any more," says Garrison. He is not crazy enough to hazard a guess as to when that might be, but according to IDC, revenues from IA-32 processors in the third quarter of 2003 were $4.637 billion and were for the first time larger than RISC-based servers, which accounted for $4.531 billion in sales. It took Intel the better part of a decade to make that happen. Itanium servers accounted for just $123 million in sales in Q3 2003 according to IDC, so as far as money goes, Itanium has a long way to go.
But in terms of shipments, Intel is clearly gunning for the RISC market. Garrison says that the volume of Itanium processors shipped in 2004 is expected to exceed the combined sales of the Alpha and PA-RISC processors from Hewlett-Packard; Itanium shipments are expected to best those of IBM's Power family of processors in 2005 and are expected to beat the RISC volume leader, Sun Microsystems' UltraSparc in 2006. "There is an absolute tipping point with Itanium, and we believe that we are very close to that point," said Garrison. "We want to make it the inevitable solution for applications." And while he didn't say this, Intel specifically is interested in replacing RISC/Unix servers with Itanium servers running Windows, Linux, or HP-UX, just like partner HP is interested in this very same thing.