At its Analyst Day event today, AMD revealed its product roadmaps and goals for 2010 and 2011. Given the depth and scope of the eight-hour briefing, we've broken the data into two posts. In Part I, we'll cover AMD's competitive strategy and CPU/GPU roadmaps as currently envisioned. Part II will examine GlobalFoundries' performance and that company's future plans.The Future Is (Still) Fusion:
One of the major themes of the event was AMD's concept of heterogeneous computing. In heterogeneous computing, workloads are divided between the CPU and GPU, regardless of whether the graphics processor is on-die, integrated into the motherboard, or a discrete card. The process of splitting and directing a given workload to the processor best-equipped to execute it is presumed to be intelligent in order to ensure GPU-centric tasks don't end up being run on the CPU or vice versa.
According to corporate vice-president Chekib Akrout, the amount of additional performance that can be achieved by focusing on single-threaded or multi-threaded performance is quite low. Heterogeneous computing on the other hand, has only begun to ramp, making it the logical focus for AMD going forward. While the term "Fusion" is often used to refer to a processor with a GPU on-package, AMD sees it as an overarching term for CPU+GPU execution. In the diagram below, APU stands for accelerated processing unit and represents the combined capabilities of the two separate processors.
Server, Desktop, and Mobile Updates
For all AMD's discussion of heterogeneous computing today, however, its actual execution has lagged considerably when compared to NVIDIA. Where the latter has sunk millions of dollars into promoting the idea of the GPU as a processing unit, AMD has generally sat back and focused on the more traditional appeal of improved 3D graphics and frame rates. There's nothing stopping AMD from turning its focus towards GPGPU development, but there's little evidence that this has been a top priority to date.
The graph above represents AMD's current server performance and the projected performance of both Magny-Cours and Interlagos. Magny-Cours is an Istanbul derivative presumably built on 45nm. The new processor will use AMD's upcoming Maranello platform, which sports four memory channels instead of two. When Interlagos launches in 2011, it'll be based on AMD's all-new "Bulldozer" architecture. Core counts are also going up—Magny-Cours is an 8-12 core processor, while Interlagos will feature 12-16 cores. In both cases, hexa-core and quad-core derivatives of these CPUs will be available for the 1P and 2P markets.
There aren't a whole lot of surprises on AMD's desktop roadmap for 2010. As we've previously reported, the hexa-core Thuban processor (aka Istanbul) will drop into the enthusiast market accompanied by a new chipset. Mainstream systems will continue using dual-core or quad-core Athlon II's, but should shift over to DDR3 by the end of the year. Dorado's integrated GPU won't be DX11-capable—we won't see that feature until 2011—but should be at least as fast as the current 785G. In 2011, AMD will roll out Bulldozer-based quad and octal cores in the enthusiast market. The mainstream segment will feature Llano, AMD's first CPU+GPU hybrid. Llano will be built on 32nm technology, and the integrated GPU will apparently be DX11-capable.
AMD has long struggled in the mobile segment, but the company's 2010 roadmap is potentially strong enough to change its fortunes. AMD will migrate to 45nm-derived mobile processors across all mobile markets next year, while simultaneously shifting to DDR3. Mobile DX11 GPUs will be available within the next 12 months, as will quad-core "Champlain"-class processors. Sunnyvale is scarcely guaranteed more market share in the next year, especially given Intel's shift to 32nm processors, but the company should see an uptick thanks to 45nm mobile Turion's presumably lower power consumption and better performance-per-watt.Conclusion
AMD's roadmap is currently an odd mixture of solid predictions and relative unknowns. Over the next twelve months, AMD generally plans to accelerate/improve already proven technology across both its CPU and GPU product divisions. In 2011, the emphasis is decidedly different. In the past, AMD has generally introduced new architectures in the server market first, with desktop parts launching 6-9 months later and mobile parts appearing last. The roadmaps AMD released today indicate that Bulldozer and Bobcat—the company's two all-new architectures—may launch across all product segments over a much shorter period of time.
If the two designs ramp as quickly as indicated, AMD's competitive position vs. Intel could shift dramatically within a relatively short period of time. Heading into 2010, AMD is generally better positioned than it was a year ago, but the company still has precious little room for error. For the next twelve months, at least, Intel looks to hold the performance high ground; whether or not AMD can change that farther on will depend on just how good Bobcat and Bulldozer turn out to be.