AMD In 2013 And New Server Plans
The company’s plans for next year are, perhaps more telling of its new direction. In 2012, AMD’s products are coming from multiple fabs employing three different nodes: 40, 32, and 28 nm. The following year, everything is at 28 nm. Chasing down the latest process (and assuming the associated risks, which sometimes result in delays) will no longer be a priority.
Instead, the emphasis is going to be on integration. To be frank, integration as a concept isn't a revolutionary idea from AMD or any other semiconductor vendor. It’s the same direction AMD’s competition is going, enabled in large part by more advanced manufacturing. Besides, remember that AMD was the company that shocked enthusiasts by putting a memory controller on its CPU die, propelling the Athlon 64 out in front of Intel’s Pentium 4. We've seen integration in action over many, many years. The difference moving forward will be an focus on the reuse of core IP, allowing AMD to put its technologies (and technologies from third parties, even) into a great range of devices more quickly.
That’s not to take away from what AMD has planned. The pursuit of integration will see AMD incorporate what is currently called the Fusion Controller Hub into its APUs, creating single-chip platforms. This is a significant development because it allows certain power and layout advantages, particularly in space-constrained form factors where typical two- or three-chip PC platforms cannot go.
Trinity will be replaced by Kaveri, which incorporates AMD’s Steamroller architecture (the second generation-evolution after Bulldozer, which AMD has already told us will incorporate notable IPC- and power consumption-oriented improvements). As with today’s APUs, those based on Kaveri should sport two or four cores. Further, AMD plans to replace its older VLIW4 architecture with Graphics Core Next, giving those upcoming APUs significant advantages in general-purpose compute applications optimized for its Heterogeneous Systems Architecture.
Oh, that’s right. It’s dropping the Fusion brand in favor of HSA. That name is making its debut today, too. The Sea Islands discrete graphics family, Kaveri APU, and Kabini APU will all include specific HSA-oriented feature that, first, allow the GPU to access CPU memory and then unify CPU and GPU memory.
The Kabini platform is expected to replace Brazos 2.0, leveraging improvements to the low-power Bobcat architecture aptly called Jaguar. It, too, gives up the VLIW4 design found in Radeon HD 6900-series graphics cards in favor of GCN.
Hondo will give way to Temash in the ultra-low power space, and that should also feature the x86-based Jaguar architecture. While Kaveri and Kabini both make appearances in AMD’s desktop and mobile roadmaps as two- and four-core configurations, Tamesh is a mobile-only concept that’ll employ two cores.
Staying Competitive In Servers
In talking with AMD’s John Fruehe, director of product marketing for AMD’s server/workstation products, it was mentioned that company plans to tackle the cloud infrastructure most specifically. Asked what it meant to target the cloud with a processor, John talked about AMD’s ability to set custom TDPs to fine-tune power consumption to maximize performance in any given power budget.
Now, I’ve never heard Intel use the term configurable TDP before, but I do know the company’s Intelligent Power Node Manager software is supposed to do something very similar, capping power consumption on a per-server basis to granularly control peak power use. How AMD plans to use this functionality in a unique way remains to be seen—though there’s no question it’s an important step forward in giving IT managers control over quickly-growing data centers with finite power.
Also interesting is that the Terramar and Sepang designs, planned to include 10 and five Bulldozer modules, respectively, were canceled. AMD says that its upcoming Piledriver architecture revision is effective enough to give it the performance it wanted from those 20- and 10-core configurations using one less Bulldozer module per die. As a result, Abu Dhabi and Seoul stand in as replacements. AMD won’t really say whether or not Terramar and Sepang would have required a new interface to accommodate their more complex layouts. However, it’s happy to assure us that Abu Dhabi and Seoul will drop into the Socket G34 and Socket C32 interfaces once they’re ready.
And with that, I’m off to Sunnyvale. Fifteen minutes turned into two hours. Looks like I’ll need to engage the afterburners up California’s I-5.