Opteron Learning to Walk, Ready to Run

by Timothy Prickett Morgan

The 64-bit, X86-compatible Opteron processor turned one year old last week, and Advanced Micro Devices hosted a big shindig to celebrate at the Rainbow Room in New York. I was too busy doing a hundred other things to goof off and go to the party, but I have spent some time with AMD in recent weeks to put together a picture of where the Opteron is going in the server market. AMD and its "Hammer" chips have come a long way in a year.

Having secured IBM one year ago as a tier-one Opteron system supplier, followed by Sun Microsystems in February 2004 and Hewlett-Packard in March 2004, you might think AMD would relax a little and talk in great detail about its plans for the server market. Not so. The company's official roadmap doesn't say exactly what AMD is up to, beyond plotting out code names and chip fabrication processes for server and workstation Opteron "SledgeHammer" chips and the related desktop and mobile versions of Athlon 64 "ClawHammer" processors. There's a lot more going on besides code names, of course. AMD has plenty of tricks up its sleeves. It is just being cagey about making promises because, if it slips, Intel and its partners and competitors will come down on it like a ton of unshipped servers.

On the server front, AMD is right now shipping two processors: the 32-bit Athlon MP and the hybrid 32-bit/64-bit Opterons. The Opterons get all of the attention, but the Athlon MPs probably represent a very large portion of server shipments. There are also probably a large number of white-box Athlon uniprocessor machines being used as servers, even though no one is necessarily counting them that way at the big IT research companies. The fact that the Athlon and Athlon MP did not take the server market by storm, wresting a big piece of market share away from Intel, is the result that the Athlons were only a little better than the Pentium IIIs and early Pentium 4s they competed against. With the Opterons, AMD has created a chip family that is arguably a lot better than the Pentium 4 Xeon DPs and Xeon MPs. And that is why Opteron sales doubled each quarter in 2003 and the company's partners shipped some 35,000 Opteron-based servers (and about 150,000 Opteron processors) for all of 2003, according to researchers at IDC. If AMD and its partners can keep doubling every quarter throughout 2004, which seems likely as 64-bit versions of Linux, then Solaris, then Windows all come live on the box through the spring and summer, then Opterons will comprise somewhere between 500,000 and 600,000 of server shipments. This will be a significant piece of the 4 million or so server shipments that are likely in 2004. At best, it seems that the absolute top end of Opteron-based server shipments would be in the 300,000 per quarter range, which would give it about 30 percent market share. If AMD does this, rival Intel can kiss its precious profits goodbye.

With the advent of low-power Opterons, which drop the voltage on chips but keep the performance of a regular Opteron, AMD has a significant advantage that Intel cannot even address. And the fact that an Opteron will do more work in 32-bit mode (and very likely in 64-bit mode) compared to an Intel Xeon for less money, at a lower clock speed, and with less electricity consumed and heat pumped out gives AMD the kinds of advantages it needs to compete with--and beat--Intel in the server space.

That is not to disparage the Athlon MPs. These were fine chips that were largely ignored by the server makers because they didn't offer enough differentiation from Pentium chips to make it worth the while and the wrath of Intel. The Athlon MPs were the final implementation of the 32-bit Athlon chips that AMD created to try to break into the server market. The current Athlon MP processors use the "Thoroughbred" core and are implemented using a 130 nanometer process at AMD's plant in Dresden, Germany. The Athlon MP has 128KB of on-chip L1 cache and 256 KB of on-chip L2 cache memory. The Athlon MP clock speeds run from 1.7 GHz to 2.1 GHz. Last summer, AMD announced a 2.1 GHz Athlon MP with 512 MB of L2 cache. Brent Kirby, AMD's product manager for its server and workstations group, says that AMD will continue to make and ship Athlon MPs as long as the market demands it. It is hard to imagine that the market will demand it for much longer, particularly as tier one and tier two server makers start concentrating on Opteron machines. AMD is already selling slightly more Opterons than Athlon MPs, in fact.

The Opterons are available for uniprocessor, two-way, and four-way/eight-way configurations. IBM and Sun are only selling two-way machines, and HP, as we told you last week, is only now reading its four-way Opteron machine after launching its two-way box earlier in the year. HP is working on a blade server using the Opterons, and IBM probably will launch one as well. Sun's blade servers already use the Athlon chip, and future machines will almost assuredly use the low-power Opterons. Sun has made the biggest commitment to Opteron, with plans to create its own server designs (created by Sun founder Andy Bechtolsheim, who just returned to Sun after the company bought his secretive Kealia Opteron server skunkworks in February) and support its own Solaris 10 operating system as a full-fledged alternative to Solaris 10 on its own Sparc processors. Sun has been vague about its plans, but it will likely deliver a range of machines that span from low-cost uniprocessor Opteron servers to fairly expensive eight-way midrange Opteron servers. As Sun does this, it will coerce others into doing so or defending with Xeon-64 alternatives.

AMD is not expected to make any Opteron chip announcements that relate to servers or workstations in the first half of 2004, which means companies will have to make do with the 1.4 GHz to 2.2 GHz Opterons and their 1 MB of integrated L3 cache memory. But according to Kirby, the company is planning to take the SledgeHammer chips and tweak them with a 90 nanometer copper/SOI process in the second half of this year. These so-called SledgeHammer-II processors will plug into the same 940-pin slots, and according to various industry rumblings, could include more on-chip cache memory. The smaller circuitry enabled by the move to 90 nanometers should allow AMD to roughly double the clock speed of the chips to 3 GHz to 4 GHz over the life of the SledgeHammer-II generation and boost the L3 cache to 2 MB or larger per chip.

The SledgeHammer-II processors will come in three flavors, just like the current Opterons. The "Venus" version of these chips will be branded as the 100 Series and will be available in regular and low-power versions for uniprocessor machines. The "Troy" version of this chip will be sold as the 200 Series and will be designed for two-way capable servers; it will also be available in regular and low-power versions. The "Athens" variant is a future 800 Series SledgeHammer-II chip, which means it can be used in four-way and eight-way machines. All of these chips will have enhancements to the Opteron core, says Kirby. But he will not say exactly what they are. HyperThreading is probably not going to be one of them, since he slammed the idea. "We don't see a big advantage with HyperThreading," he said. "We would do a multiple core chip if we did anything like that."

And while Kirby would neither confirm or deny this, the future SledgeHammer-III processors, expected in the second half of 2005, are rumored to be the first generation of AMD chips that will sport two Hammer cores on a single chip. This is about the same time that Intel will deliver the dual-core "Tulsa" Xeon-64 processors. The third generation Opterons go by the code names of "Denmark" (100 Series), "Italy" (200 Series), and "Egypt" (800 Series). They will be implemented in a 90 nanometer copper/SOI process, just like the SledgeHammer-IIs.