The Four Hundred

Intel two weeks ago announced its third generation of high-end 64-bit Itanium processors, the so-called "Madison" chip, and also debuted faster "Gallatin" Pentium 4 Xeon MP processors that are used in four-way and larger Intel servers. While neither the Madison nor the Gallatin processors will be used inside iSeries servers themselves, somewhere between 75 and 80 percent of OS/400 shops also have Intel-based machines, so what Intel is up to does affect them.

With the Madison Itanium 2 processors, Intel seems to be finally hitting its stride in the 64-bit world, which has been dominated by the remaining RISC/Unix vendors--IBM, Hewlett-Packard, Sun Microsystems, and Fujitsu Siemens. The third time had better be the charm for Intel, which has pumped an estimated $5 billion of chip design and manufacturing technology into Itanium and has yet to get any substantial revenue stream from the processors.

While OEMs have made either serious or half-hearted commitments to Itanium, HP has staked the future of its company on the processor. With Madison, it seems like Itanium is a much safer bet than it was three years ago.

"This is like a marathon," explained Mike Fister, senior vice president and general manager of Intel's Enterprise Products Group, which makes server and workstation chips as well as chipsets. "People tell me that Itanium has only a fraction of the sales of Xeons. And while this is true, it would be wrong to think that we do not have momentum."

If Fister sounds a little defensive, he is, as are other executives at Intel who have backed Itanium. But Intel has a bit of spring in its step these days, thanks to Madison. It's a good chip, even if it isn't perfectly compatible with the venerable X86 architecture that made Intel a household name.

The Madison will be sold under the Itanium 2 brand name, just like its predecessor, the "McKinley" second-generation Itanium. The Madison chip runs at 1.3 GHz and 1.5 GHz, compared with the 900 MHz and 1 GHz clock speeds of the McKinley processors. The McKinley chips offered 1.5 MB or 3 MB of on-chip L3 cache memory, had about 250 million transistors (with 3 MB of cache), and were implemented in a 180 nanometer process. The Madisons, which are implemented in a much tighter 130 nanometer process, allow Intel to put a lot more L3 cache on the chip and to crank up the cycle time on the processors without causing the whole shebang to melt and release that magical blue smoke, which is the essence of computing. The Madison chip has a whopping 400 million transistors, according to Fister, and a lot of that is taken up by the large L3 caches, which can significantly boost the performance of commercial applications, particularly those that are sensitive to cache, like those written in Java.

Intel has three different versions of the Madison chip, each with a different price and performance point. The entry Madison chip runs at 1.3 GHz and has 3 MB of L3 cache. It will sell for $1,338 in 1,000 unit quantities. The middle chip is slightly faster, at 1.4 GHz, and has slightly more L3 cache, at 4 MB; it costs $2,247 per 1,000. The fastest Madison chip runs at 1.5 GHz, has 6 MB of L3 cache, and costs $4,426 per 1,000. These are the same prices Intel has been charging for its 900 MHz/1.5 MB L3 cache, 1 GHz/1.5 MB L3 cache, and 1 GHz/3 MB L3 cache McKinley chips. OEMs and customers expecting a big price break on the Itanium processors with the Madison launch are going to be disappointed.

But, with anywhere from 40 to 50 percent more oomph, even at the same prices that is still a 40 to 50 percent improvement in price/performance for the raw processors. Whether this translates into price/performance gains, compared with McKinley-based machines, is largely irrelevant, since there were so few McKinley machines sold anyway and only a miniscule number of first-generation "Merced" Itanium machines ever made it out of the OEM factories. The Madison Itaniums are all about comparisons with existing and entrenched RISC/Unix architectures. That's why Intel has been letting OEMs like HP and NEC release performance information about Madison ahead of the launch date. It wanted to give Unix customers who buy machines with IBM Power, Sun Sparc, and even HP Alpha and PA-RISC processors, the time to think about it before they made an acquisition.

Given the performance numbers that Intel has been showing, the availability of Windows 2003, HP-UX, and Linux on the processors, and the fact that 400 key ISV applications are now available on Itanium (Fister says this is up from 100 applications this time last year), it looks like Itanium might have finally hit critical mass, so it can start getting some traction in commercial computing. Fister said that there are 44 tools for developing applications on Itanium for Windows, 12 tools for doing so on HP-UX, and 30 tools for doing so in Linux. Intel also has six different language compilers that it has created for Itanium.

Like Windows NT 4.0 was the third-time charm for Microsoft in the enterprise, it looks like Madison might just turn out to be the third-time charm for Intel in the enterprise. A lot will depend on how the OEM workstation and server makers try to peddle or meddle with Itanium products. While Fister can say that the number of OEM Itanium configurations with two or four processors has grown from 20 to 40 in the past year, and the number with eight or more processors has grown from five to 10, the enthusiasm with which these key customers of Intel push Itanium will determine, more than any feeds and speeds, how well or how poorly Itanium takes off.

If the Itanium 2 chip gets enough software support, IBM could take a single processor--perhaps in the "Deerfield" generation next year, but more likely much later--and put it into the iSeries Integrated xSeries Server (IxS) coprocessor that it sells to plug into the iSeries bus. IBM would have to have a compelling reason to make an Itanium-based IxS--perhaps to support decision support, data warehousing, or other numerically intensive workloads. It seems far more likely that IBM will point customers who need the processing and memory capacities of the Itanium 2 to its xSeries 450 machine, which will start shipping at the end of July, and it is even more likely IBM will point such customers to its 32-bit Intel-based servers or its 64-bit pSeries AIX and Linux servers for this kind of work.

While IBM doesn't officially support the xSeries 450 with the Integrated xSeries Adapter (IxA) card, for clustering external xSeries machines to iSeries boxes, it could probably do so easily if it were motivated. For most iSeries customers, the existing two- and four-way xSeries machines using Xeon processors will provide enough oomph for these kinds of workloads. The xSeries 450 was launched in May, just as a bug was found in some of the McKinley predecessors to Madison, and IBM halted shipments until Madison became available.

The really useful thing about Madison is that the advent of this processor in midrange and enterprise iron is putting pricing pressure on all server lines, including the iSeries. While Madison is not much in terms of sales, those vendors who want to sell other midrange and enterprise machines have to keep pace with it or they will inadvertently foment the Itanium market they now mock.

The speed bump on the Gallatin Xeon MP processors is probably more relevant to OS/400 shops than the Madison announcement. Gallatins are used in more than 60 different 32-bit servers with four or more processors. Odds are, if an OS/400 shop is buying a four-way Intel-based server today from one of Intel's OEMs, it is buying a Gallatin machine.

The performance improvements from the faster Gallatin chips will be particularly interesting to those midrange customers who are running hybrid OS/400-Windows applications. Many big ERP, SCM, or CRM installations these days have OS/400 supporting a central database but Windows servers running the applications themselves. It is easier to manage one or a few fast application servers than it is to cope with many slower application servers.

The performance improvement in the new Gallatin chips will, ironically, make it somewhat more difficult to justify the move to the Itanium processors. Before the announcements two weeks ago, Gallatin chips topped out at a 2 GHz clock speed. On many commercial applications (like the TPC-C online transaction processing benchmark test), chip for chip, a 2 GHz Gallatin matches the performance of a 1 GHz McKinley Itanium 2 processor. With Intel ratcheting up the clock speed of the Gallatin chip to 2.8 GHz on June 30, the only real performance benefit that the new 1.5 GHz Madison Itaniums will offer over new Gallatins will be on jobs that require more than the limit of 4 GB of main memory per processor, which that 32-bitness of Pentium processors impose. If Intel stopped making Gallatins, there would be a lot more interest in Madison. And there would probably be a lot more interest in the "Hammer" 64-bit Opteron processors, made by rival Advanced Micro Devices, which explains in part why Intel continues to improve its 32-bit server processors.

The first batch of Gallatin chips came in at 1.5 GHz/1 MB L3 cache, 1.9 GHz/1 MB L3 cache, and 2 GHz/2 MB L3 cache configurations. They sold for $1,177, $1,980, and $3,692 respectively, in 1,000-unit quantities. While the new cranked Gallatins will plug in to the same sockets as the old ones, some OEM vendors (including IBM) have refreshed their server lines to coincide with the new Gallatins and are encouraging customers to upgrade their servers in conjunction with moving to the new, faster processors. (See the July 9 issue of Guild Companies, Windows & Linux Edition, for details on the new Itanium and Gallatin servers from Intel's OEMs.)

The new Gallatins are based on the same 130 nanometer chip technology, and they come with either 1 MB or 2 MB of L3 cache on chip. There are three clock speeds: 2 GHz, 2.5 GHz, and 2.8 GHz. As with past server chips from Intel, only the fastest processor in the new line gets the largest on-chip cache. As was the case with the McKinley-to-Madison jump in the Itanium line, Intel is holding prices constant on the new Gallatins versus the older ones. The performance improvement on many commercial workloads normally scales with clock speed, as long as some increased L2 or L3 or cache is added. (Intel has not boosted the cache sizes on the faster Gallatins.) With the caches held constant on the Gallatins, the performance improvement should be in the range of 20 to 40 percent, and less on cache-sensitive workloads.

Intel knows that many workloads will require a larger L3 cache, which is why it will debut a Gallatin processor with 4 MB of L3 cache early next year, and the future "Potomac" version of the Pentium Xeon MP processor, due in late 2004 or early 2005, will also feature increased L3 caches. Potomac will be the first multiprocessor server chip that Intel creates using 90 nanometer and 300mm wafer chip-making technologies.