AMD Discusses Details of Next-Generation Products

At Hot Chips 22 today, AMD will be discussing new details of its Bobcat and Bulldozer processors—but don't be surprised if some of the new information sounds familiar. With the exception of a few tidbits on Bobcat's architecture and a new Bulldozer block diagram, much of what the company discussed with us recently was revealed last November at the company's Analyst Day.

For those of you who might not keep up with the company's roadmap, here's the current situation. Llano, the 32nm Phenom II-derived APU (read: CPU+GPU) that AMD was hoping to launch at the end of Q4 this year won't tip up until Q1. Normally this means "March 31st," but AMD implied that it might launch the chip earlier in the quarter rather than later. Llano will target the low-end and midrange desktop segment. AMD will also launch its new Ontario mobile processor late this year. Ontario combines the Bobcat mobile processor with an integrated GPU; the term Bobcat refers strictly to the CPU core. AMD is billing Bobcat as a "sub one-watt capable core" (emphasis ours), and predicts that it will clean Atom's clock while redefining AMD's position in the netbook / low-end notebook space. Unlike Llano, which was (unofficially) pushed back, Bobcat has actually been pulled forward—AMD's yields have been sufficiently good to merit an early launch it seems.

Finally, there's Bulldozer. This is AMD's new high-end server / workstation / enthusiast CPU. Like Bobcat, it's an all-new architecture from the ground up. AMD is steadfastly refusing to give any specific timeframe for when Bulldozer chips might be available. But Bulldozer, when it arrives, will be drop-in compatible with the current generation of Opteron motherboards as well as AM3 products, provided that your motherboard manufacturer provides the necessary BIOS update.

Now that we've covered those basics, let's examine the two new cores.

Big Bad Bulldozer



Here's Bulldozer as we've seen it for the past nine months. Bulldozer is designed to efficiently combine certain execution units and capabilities when possible without compromising performance. Each Bulldozer module is a mixture of shared and dedicated components, like so:



Each Bulldozer module shares a common front end, floating-point unit, and L2 cache, while the L3 cache is split between all the modules on a die. According to AMD, using two independent integer units allowed for a dramatic performance boost per module, while combining other features reduced die size, power consumption, and complexity. AMD told us it arrived at the above model after an enormous amount of analysis during which engineers studied the activity level of every core-level component in a variety of workloads. FPU performance should be at least equal to Magny-Cours in identical code.



Here's a block-level diagram of an eight-core / four module Bulldozer chip; we expect Bulldozer to launch in this configuration, among others. Bulldozer supports all modern SIMD sets, including Intel's next-generation AVX. Before Intel announced AVX, AMD had announced its own SSE5 SIMD, but has since deprecated that design in favor of adopting Intel's standard.

The company did not to provide any information on cache size, clockspeed, or launch window. As for performance data, the company said, "AMD delivers 33 percent more cores and an estimated 50 percent increase in throughput in the same power envelope as Magny-Cours." This is slippery comment, since we don't know the precise definition of the terms "power envelope" or  "throughput"--at least not in this context--and not without some information on what benchmark was used for the comparison. Broken down, according to this quote, AMD predicts that a Bulldozer core will have roughly 12 percent higher "throughput" than a Magny-Cours core—whatever that means.

And Now For Something Completely Different:
Next up, there's Bobcat:



Bobcat is arguably the more interesting of the two processors. As with Bulldozer, AMD has covered all of this information before, but just in case you've missed it:  Bobcat targets an operating TDP of 1-10W, includes support for SSE1-SSE3, is an out-of-order execution processor (which should make it considerably more efficient, clock-for-clock, than Intel's Atom) and it'll launch late this year as a part of AMD's "Brazos" dual-core notebook / netbook platform. Ontario is AMD's first move into netbook territory and could redefine the ultrathin category of laptops AMD has been promoting for several years.

As you'd expect, this new mobile chip has been extensively optimized for low power consumption and efficient operation; AMD claims Ontario will deliver "90 percent of today's mainstream performance in less than half the silicon area." Again, this is the sort of PR statement that tells very little. AMD didn't quantify "mainstream performance," and we've been provided with a comparator of secondary importance. A small die is important, for multiple reasons of course, but we'd much rather read that Ontario "delivers 90 percent of today's mainstream performance at 50 percent of the power consumption."

Other than the above, the only other thing Sunnyvale admitted was that Ontario should beat Atom silly. The only bits of genuinely new information on Bobcat, as far as we can tell, are the size of the chip's L1 caches (32K each, 64K total) and the fact that it supports secure virtualization. When we asked about the chip's onboard GPU and its capabilities, AMD refused to comment.
Conclusion:
We're quite excited over Bobcat—while the chip won't compete with Atom at the UMPC level, Ontario could strike a new balance between power consumption and performance. AMD may not make much money on Bobcat sales—not at first—but the new core architecture could help Sunnyvale break Intel's decade-long stranglehold on the mobile industry.

We don't feel quite as confident about Bulldozer. Unlike Bobcat, which opens an entirely new market for AMD, Bulldozer is tasked with reversing the company's losses in the server, workstation, and high-end enthusiast markets. Even if Bulldozer were to launch as early as Q1 of next year, it'd be taking on Intel's next-generation Sandy Bridge—and there's zero sign that the core will drop anywhere near that soon. To date, AMD has fought back in the server market with deep price cuts and by offering a higher number of cores per dollar.

The longer Bulldozer is delayed, the greater the chance that it'll debut into the teeth of new six-core and eight-core Sandy Bridge products. To date, AMD's strategy of aggressive pricing and a higher number of actual cores has served the company well, but if its brand-new architecture can't match Sandy Bridge chips clock-for-clock, AMD will be in a similar situation as they are today.
If you liked this, be sure to check out our architecture deep-dive HERE.

Via:  AMD
blog comments powered by Disqus