AMD Blazes New Path with Bulldozer
Now that AMD has jettisoned its chip production business with the Globalfoundries spinoff, it can concentrate on what it has always done best: microprocessor design. Much of its success early in the decade resulted from outmaneuvering Intel, its much larger rival, in the lucrative x86 server space. With the Opteron CPU, AMD paved the way for the next-generation x86 platform with 64-bit processing, integrated memory controllers, and a NUMA architecture. Now with Bulldozer, AMD’s upcoming x86 core, the chip vendor is once again looking to leapfrog the competition.
Bulldozer represents AMD’s first new x86 core redesign in seven years, according to Dina McKinney, vice president of design engineering at AMD. McKinney briefed reporters and analysts last week on the new architecture, in preparation for a more public unveiling at this week’s Hot Chips conference at Stanford University. The intention, says McKinney, is for this core to “live for a long time.”
AMD is actually talking up its two new core architectures this week, Bobcat and Bulldozer. Bobcat is AMD’s low-power core and is initially being targeted to notebooks and netbooks, where it will compete against Intel’s Atom processor. The Bulldozer architecture will be the basis for Opteron chips and high-end desktop CPUs, where performance and scalability are paramount.
The focus of the Bulldozer design is to optimize thread throughput against die real estate and power consumption. Intel has attacked the issue with HyperThreading, its version of Simultaneous MultiThreading (SMT), where each core can handle two threads with minimal hardware redundancy. Meanwhile, AMD has stuck with the one core per thread model, known as the Chip MultiProcessing (CMP).
In Bulldozer, the company has opted for a sort of hybrid approach where each module consists of two integer schedulers, which appear to the software as two separate cores. Because they appear as individual cores, AMD is counting them as such. For example, a 4-module Bulldozer CPU would be sold as an 8-core processor.
The dual integer schedulers shared a floating point unit, which consists of an FP scheduler that manages two 128-bit multiply and accumulate units. An integer unit can also use the two 128-bit FPU units to schedule 256-bit operations, when extra wide floating point computations are called for. Each Bulldozer integer unit has is own L1 data cache, but they, along with the floating point unit, share an L2 cache. To build a processor, multiple Bulldozer modules are laid out on the chip and they all share an L3 cache, along with an integrated memory controller and a Northbridge controller.
The whole idea is to strike a balance between dedicated and shared hardware such that those resources most in demand (the integer unit and L1 cache) are duplicated and those in lesser demand (FP unit and L2 cache) are not. (Note that even in technical HPC applications, the integer unit dominates execution cycles.) AMD’s claim is that the Bulldozer design delivers 33 percent more cores and an estimated 50 percent increase in throughput within the same power envelope as the current generation Magny-Cours Opteron.
The first Bulldozer-based CPU will be the Opteron 6000 series “Interlagos” CPU aimed at enterprise and HPC server platforms. Interlagos is a 16-core chip — thus it will be built from 8 Bulldozer modules — and will be the first AMD processor to use the 32nm SOI manufacturing technology.
Interlagos is scheduled to go into production sometime in 2011. It will use the same G34 socket as its Magny-Cours predecessor, so even though the underlying microarchitecture has changed, AMD is promising a plug-in upgrade for the first Bulldozer Opterons. The 4000 series Opterons could swallow Bulldozer technology as well, but there is no such product yet announced for this line.
Bulldozer will also end up in the 8-core Zambezi processor for high-end desktop systems, again in 2011, but following the Interlagos release. McKinney hinted that Bulldozer cores would eventually make their appearance in APU-type (CPU-GPU Fusion) processors and even mobile chips, but presumably such products won’t show up until 2012 and beyond.
Whether the Bulldozer technology ignites a comeback for AMD in the server and workstation arena remains to be seen. The company is not releasing performance or pricing data on future Bulldozer-based processors yet, so there’s no way to gauge their competitiveness against Intel’s Xeon chips.
AMD has been losing market share in this space for years. According to IDC, Intel supplied a whopping 93.5 percent of server processors in the second quarter of 2010, while AMD claims a measly 6.5 percent. Even in HPC, where AMD chips traditionally have had better traction, the numbers aren’t much better. On the latest TOP500 list (June 2010) of supercomputers, Opterons have only a 9.8 percent share, compared to Xeons at 80.2 percent. Certainly AMD has lots of lost ground to make up.