AMD’s 2014 roadmap: A hazy future for desktop x86 chips as AMD doubles down on ARM

AMD’s updated roadmaps for 2014 show plenty of transitions, but not necessarily in the product lines that interest enthusiasts the most. Piledriver — the eight-core, AM3+ processor based on Bulldozer — will remain AMD’s top-end desktop chip throughout next year, even though Steamroller-based Kaveri is preparing for a January unveil. In the past, AMD has rolled out updates to server, desktop, and APU families in fairly short order, so this decision not to build server or desktop variants on the new CPU is a bit surprising.

All of this might be a moot point, if Kaveri had hit its frequency targets. Increasingly, however, this seems to not be the case. At APU13 last week, several Kaveri systems were on display with a base clock of 3.5GHz and an unknown Turbo Mode. That’s sharply below Richland, with its 4.1GHz base, 4.4GHz Turbo, and 5GHz overclock — and it raises the question of whether or not the chip will deliver the performance benefits AMD initially promised. A 3.5 to 3.8GHz clock speed would be a 15-20% reduction from Piledriver’s level — easily enough to offset the increased single-thread performance.

AMD’s server roadmap shows Piledriver continuing as a high-end part through 2014 as well, but the new ARM Cortex-A57 core from AMD takes over the budget/low-power segment by the middle of next year. Interestingly, AMD’s Kabini servers are apparently nothing more than a stopgap measure, and shuffle off the mortal coil to be replaced by ARM variants. This is rather odd, as it implies AMD is going to go after the low power market strictly with ARM, when it has an x86 core that fits into the segment as well its recently acquired SeaMicro subsidiary that specializes in x86 server hardware.

Kabini and Temash transition to Beema and Mullins next year, with minimal changes. The thing is, the current slide is wrong. AMD already offers Kabini parts in a 10W envelope, as we covered at the chip’s debut. The Temash to Mullins debut, with a 2W envelope, does imply that the follow-up chip will slip into a better space. Overall, however, the best way to think of this update is in line with what we saw between Trinity and Richland, or Brazos and Brazos 2.0 — a modest improvement on some key metrics, but not a fundamental game-changer.

Finally, in embedded, we see the ARM Cortex-A57 cores debut as higher performance options than Kabini, but below Steamroller. This is interesting, for several reasons. First, we know Kabini and Temash are quite competitive with Bay Trail, which is generally as fast as Apple’s A7 in thermally constrained environments. This implies that the A57 hits an even higher performance target, or at least that AMD thinks it has the potential to do so. We also see Hierofalcon hitting 4-8 core deployments with up to 30W of dissipation, which stretches the traditional definition of embedded. Steppe Eagle, the Kabini variant, is aiming at a 5W-25W envelope, but still contains a GPU, while AMD’s first A57 core is CPU-only.

Tahiti retrenches, holds its ground in 2014

In its earnings call last month, AMD announced that it would tape out 20nm GPU designs in the first part of 2014. That surprised us — with 20nm production supposedly starting in the early part of the year, we expected AMD to have designs ready to ship fairly early in the ramp. GPUs are typically early production candidates; AMD and Nvidia have quickly moved to new nodes as they came online. If AMD is taping out designs in Q1 2014, it suggests we won’t see new cards until Q1 2015. While we’ve heard rumors of Nvidia’s Maxwell being pushed back into later in 2014, there’s nothing to suggest it’s actually running a full year behind.

What’s interesting about this slide is that it does suggest that AMD is adjusting its product placement. The R9 290X and R9 290 don’t change, but the Radeon 8800 family, for example, corresponds to the Radeon 7800 series — meaning those GPUs have 1,280 cores. The R9 280X that AMD shows as replacing the Radeon 8800, in contrast, is the full implementation of Tahiti, with a full 2048 cores. These changes ripple down the entire product family.

Keep in mind, we have no idea when the repositioning will happen, so I’m not suggesting skipping a GPU purchase if you’re in the market today, but this slide suggests that AMD is already anticipating changes it will make to remain competitive against Nvidia, should Team Green have new hardware in the pipeline.

An uncertain future for AMD’s “Big Core” business

What we’re seeing here, I think, is the inevitable result of a resource-constrained company attempting to fight a multi-front war. In the past 18 months, AMD ramped two new APUs (Kabini, Kaveri), built SoCs for both the Xbox One and PS4, built a new high-end variant of its 28nm GPU architecture, overhauled the Radeon HD 7000 driver family, and built a new ARM processor core. That’s an enormous amount of work, particularly for a company struggling back towards profitability.

The good news on AMD’s roadmap comes courtesy of Kaveri’s new GCN architecture and HSA implementation, the first ARM core, and a refreshed Temash APU that fits into a 2W envelope. The more troublesome news is the lack of a CPU refresh for server and enthusiast products where it would be useful, combined with troubling rumors of Kaveri’s CPU performance. We’ve heard rumors that AMD temporarily suspended work on all of its Steamroller follow-ups to bring the Xbox One and PS4 to market in time — and that would explain why we don’t see larger Steamrollers, a platform refresh for Socket AM3, or much news on the big core server front.