AMD Bulldozer Chip Wants To Flatten Intel

Bulldozer Opteron 6200 processor, first x86 chip to use up to 16 cores, aims to help servers host many virtual machines while consuming less power per core.

AMD launched its Bulldozer Opteron 6200 processor Monday, the first x86 chip to be manufactured with up to 16 cores. The previous Opteron 6100 maxed out at 12 cores.

The more cores, the better the chip tends to be at hosting virtual machines. Bulldozer represents a redesign of the original, eight-year-old chip; the new Opteron 6200 has been optimized for both virtualized workloads, with a low-power variant, the Opteron 4200EE, also optimized for lower power consumption and reduced-cost cloud operations.

At the same time, it's something of a gamble for AMD, which needs a winner in the server chip market to revive its languishing fortunes. In launching Bulldozer, AMD is unveiling a new micro architecture that alters the definition of a processor core.

The original Opteron took market share from Intel's Xeon series, with server designers like Sun Microsystems' Andy Bechtolsheim snapping it up to produce new x86 servers. The Bulldozer Opteron is a complete makeover of the chip in an attempt to gain back what has become an eroding AMD share of the chip market.

"They're real workhorses, capable of running lots of virtual machines," said Margaret Lewis, director of server software for AMD, in an interview. That makes Bulldozer a candidate for cloud service providers as well as enterprise data centers that wish to maximize workloads per virtualized host, she said. For example, a single rack full of 16-core 6200s can host 672 virtual machines, with each VM having its own core, Lewis said.

At the same time, the Bulldozer architecture represents a rearrangement of chip components on the surface of the chip die. Lewis called Bulldozer's combination of two eight-core dies a chip "module." What's new is that each module represents two integer units and a single floating point unit. Under an Intel definition of cores, each core would have its own integer and a floating point unit. AMD has reshuffled the deck and announced that two cores have two integer units and a single, shared floating point unit. This makes sense since in virtualized server operations, where the floating point unit is typically used less than half the time, compared to integer operations.

Under certain circumstances, critics say this approach will hurt single-threaded performance involving floating point operations, such as in a scientific application on an individual desktop. That doesn't appear to be AMD's concern as it aims for the virtualized enterprise server and cloud data center markets.

"The integer units can work faster because they share things that make sense to share," said Lewis. That includes the Level 1 and Level 2 caches as well as the floating point unit.

That's a good idea if you're looking to increase real estate on a single chip for the parts of the processor that speed up handling the threads launched by many independent virtual machines or many cloud workloads.

AMD's Michael Detwiler, server product marketing manager, told Ars Technica that the chip adds 25% to 35% more processing power over the Opteron 6100. AMD's announcement also claimed the chip was 84% more powerful than the Intel Xeon 5600. But that's comparing the latest generation AMD chip to an earlier Intel chip with only six cores, Ars Technica warned.

The low-power variant of Bulldozer, the 4200EE, consumes just 35 watts, spread across eight cores or about 4.4 watts per core. The lowest power Intel x86 chip, AMD said in a footnote to its announcement, is the L5630 Xeon, which spreads 40 watts across four cores or 10 watts per core. (Intel's low power Atom chip is not part of the x86 family.)

In addition, Lewis said the Bulldozer chips "have knobs to run to control server power consumption." If a core goes to an idle state, it can be shut down until it's needed again, she said. And some cores can be run at full throttle or maximum power for which they were designed, while others run in power conserving mode.

Critics point out that elements of AMD's new architecture, such as its intentionally lengthened data pipeline, resemble the Pentium Pro 4 format, which turned into a spectacular Intel failure. But AMD's Lewis prefers to say that AMD has put together a combination of tradeoffs for today's server market. It is willing to take some knocks for circuit design if the Bulldozer modules yield more virtual machines served using less power. That's a combination could play well with some purchasers in today's market.

An October survey by InformationWeek on server purchasing found x86 servers were more in vogue than ever, with concerns about the heat they generated as more units were added to the data center.

February 21, 2011 by Tim Fischer

For a fuller discussion of how AMD is changing, or blurring, the definition of what constitutes a CPU core, see David Kanter's AMD's Bulldozer Microarchitecture.