AMD Rolls Out Dual-Core Opterons Early

by Timothy Prickett Morgan

Chip maker Advanced Micro Devices has stepped its competitive game in the server market against arch rival Intel as it has launched its first dual-core Opteron processors out the door a few months ahead of schedule. AMD moved up the announcements (which happened just after we went to press in the last issue) to give it more time to make some money, since Intel is not expected to launch dual-core Xeon and Itanium processors until late 2005, with shipments in actual systems not expected until early 2006.

For now, AMD is only rolling out its dual-core Opterons--the third generation of its chips and the second ones implemented in a 90 nanometer process--in the 800 Series, code-named "Egypt." The Opteron 800 Series can be used in four-way and larger systems using SMP and NUMA processor clustering, but for now, they are mostly used in four-socket servers. Rather than wait until the 100 Series (code-named "Denmark" and used in uni-socket workstations and servers) and 200 Series (code-named "Italy" and used in two-socket boxes) to be ready and do the dual-core Opteron launch as an entire family, AMD has made the smart move and did exactly what Intel cannot do with its dual-core launch. Go after servers.

Intel is planning to launch its dual-core Pentium 4 processors into the high-end PC market any day now, but it is obviously not ready to do dual-core processors in the Xeon and Itanium server space. It is fair to say that, based on the hodge-podge designs that Intel announced at its Intel Developer Forum a few months ago and the shifting Itanium roadmap for the past two years, that Intel was planning for dual-core processors quite a bit further down the road from now. I happen to think Intel thought it could push clock speeds up to the 5 GHz or 6 GHz range and that server makers would come up with ways of dissipating the heat, and only after that would it have to resort to putting multiple cores on a chip to further boost performance. But customers are balking at their electric bills and server makers do not want to have to completely redesign their server chassis every chip cycle, so now the thermal envelopes are nearly set in stone. Chip makers like Intel and AMD can do whatever they want with their transistor budgets--so long as it doesn't consume more electricity and create more heat.

In any event, while Intel is starting on the desktop with its dual core chips (something it can do because of the relatively small size of the Pentium 4 compared to the Xeon or the Itanium chips), AMD is going for the low-volume, high-profit server chip for its first dual cores, and Pat Patla, director of server and workstation marketing at AMD, says that it is making the right move and that it is not just about going where Intel isn't ready. "If Intel went dual-core first in servers, I would still go after servers, too." It all has to do with expected yields, which are always low with a new chip making process and often low with a new chip design. If you are going to have a low-yield product, you want to be able to charge as much as possible for it, which is why traditionally in the RISC/Unix server and workstation market--back when all of the server makers used to make their own chips--the midrange and often the high-end servers tended to get the latest-greatest RISC chips first, and then they would cascade down or up the line and then into workstations. AMD is taking a page out of the RISC/Unix playbook, and this is something it can do because there are three server makers--Hewlett-Packard, Sun Microsystems, and IBM--who are also pushing Opterons in new workstations as well. Intel is pushing Xeon DPs in workstations, Itanium in high-end servers (even HP has dropped its Itanium workstations), and Xeon MPs in midrange servers. In some ways, Intel and its partners have to juggle more architectures than AMD and its partners do.

But both vendors have processors--whether single- or dual-core--that are tailored to specific scalability needs--one, two, or four or more sockets. Because of the yield issue, it is smartest to start with the Opteron 800 Series because there will be far fewer customers for these processors than for the Opteron 200 Series and, potentially, the Opteron 100 Series if server makers ever get it through their heads that uni-socket machines are a big market and with dual-core chips, you need to be selling one. (If anyone understands this, it is Dell, which owns the uniprocessor server market and has been tossing and turning over the opportunity that IBM, HP, and Sun have left wide open by not shipping a single-socket AMD box. But, Dell's loyalty to Intel, which gives it plenty of advantages, is stronger than its desire to go 64-bit, dual-core rapidly in the entry server space. That is, for now.)

The dual-core Opterons taped out in June 2004 and were in demo systems in August 2004. All of the dual-core Opterons are pin-compatible with the 940-pin sockets used in the prior generation of 90 nanometer single-core Opterons. However, systems that were designed to support the first-generation of Opterons, which were created using a 130 nanometer process and have different voltages, cannot support the new dual-core chips. The voltage regulators and other electronics are completely different. So, to be specific, the IBM eServer 325 and the initial batches of Sun's V20z Opteron servers cannot use the dual-core Opterons. (There are undoubtedly lots of white box machines that cannot, either.)

For now, AMD is announcing three 800 Series dual-core parts, and they are also based on the Rev E iteration of the Opteron core. All of these chips fit into the 95 watt power envelope of the single-core Opterons and all of them have 1 MB of integrated L2 cache per core, as well as their own 64 KB data and 64 KB instruction caches. Each chip has a system request queue that is fronted by an on-chip crossbar switch that in turn is fed by the integrated DDR memory controller (with two 72-bit memory ports that deliver 6.4 GB/sec of memory bandwidth each into the crossbar using DDR-400 main memory) and three HyperTransport links (which have an aggregate of 24 GB/sec of peak bandwidth). The HyperTransport links hook into other chips directly (not through a shared front side bus) and into I/O subsystems.

The 865 chip runs at 1.8 GHz, the 870 runs at 2 GHz, and the 875 runs at 2.2 GHz. (The single-core 852 runs at 2.6 GHz, and just like other dual-core architectures, AMD has had to reduce the clock speed a little to stay in its thermal envelope and to get decent initial yields.) These chips are shipping now, and Patla says he expects server vendors to begin shipping them in systems in May. The dual-core 200 Series processors will start shipping out of the Dresden, Germany fab, in about three weeks, according to Patla and will be in systems in about late May or so. The dual-core Opteron 265, 270, and 275 processors will run at the same 1.8 GHz, 2 GHz, and 2.2 GHz clock speeds. The 100 Series will begin volume shipments in June and are expected to be in systems by July. Eventually, says Patla, AMD will roll out the low-voltage and low-power variants of these chips, the Opteron EE and HE, which have thermal envelopes of 30 watts and 55 watts, respectively.

In terms of performance, the top-end 2.2 GHz dual-core Opteron chips will have anywhere from 30 to 75 percent better performance compared to the 2.6 GHz single-core Opteron chip. Patla says that the 1.8 GHz dual-core chip offers between 20 and 30 percent better performance than that 2.6 GHz single-core processor, and they will have the same prices in the 100, 200, and 800 Series. Here's how the Opteron prices will stack up for 1,000-unit quantities:

Opteron 100 Series:

• Model 175 (2.2 GHz, dual): $999
• Model 170 (2 GHz, dual): $799
• Model 165 (1.8 GHz, dual): $637
• Model 152 (2.4 GHz): $637
• Model 150 (2.6 GHz): $417

Opteron 200 Series:

• Model 275 (2.2 GHz, dual): $1,299
• Model 270 (2 GHz): $1,051
• Model 265 (1.8 GHz): $851
• Model 252 (2.6 GHz): $851
• Model 250 (2.4 GHz): $690

Opteron 800 Series:

• Model 875 (2.2 GHz, dual): $2,649
• Model 870 (2 GHz, dual): $2,149
• Model 865 (1.8 GHz, dual): $1,514
• Model 852 (2.6 GHz) : $1,514
• Model 850 (2.4 GHz): $1,165

On the desktop, the "Toledo" chip is the dual-core variant of the Athlon-64 FX chip. The Toledo chip will be branded as the Athlon 64 X2, and it will come in four flavors: the 4200+ running at 2.2 GHz with 512 KB on-chip L2 caches; the 4400+ running at 2.2 GHz with larger 1 MB caches; the 4600+ running at 2.4 GHz with 512 KB caches; and the 4800+ running at 2.4 GHz with 1 MB caches. The Toledo chip has around 233 million transistors and will plug into the same 939-pin socket that prior 90 nanometer Athlon 64 chips used. The main difference between the Toldeo and the other Opteron chips seems to be that the Toledo chips can have smaller L2 caches and have only one HyperTransport port. AMD will begin shipping the Athlon 64 X2 processors sometime in this quarter, and they are expected to be in systems around June.