A new, unofficial roadmap from AMD has shed light on the company’s plans for the 2014-2015 timeframe. There are some specific reasons I’m uncertain if this roadmap is legitimate – but since it’s also rather interesting, let’s dive into what it shows. If accurate, AMD’s FX series of products is essentially moribund, the next mainstream CPU upgrade won’t arrive until 2015, and AMD is planning to launch a new socketed Kabini part in mainstream systems.
In fact, if this roadmap (pictured below) is accurate, it’s possible that AMD is mulling over an exit strategy for its high-end big-core x86 business. That’s an explosive claim to drop, so let me explain my reasoning before you jump on me.
First off, if this roadmap is true, the FX-9590 was essentially AMD’s swan song for the high-end. The company has no public plans for an eight-core Steamroller part for the AM3+ segment or Opteron.
This roadmap suggests that Carrizo, Kaveri’s replacement, will integrate a new CPU but keep the same GCN architecture that Kaveri uses.
That makes sense if AMD intends to tape the chip out in the next few quarters. AMD may be adopting a similar cycle to Intel, where it improves on the CPU architecture in one update, then the GPU in the following cycle.
The small core section of the roadmap – the Kabini/Beema portion – is data we’ve covered recently. We already know that these chips are tweaked versions of Jaguar with better power consumption and that they’re built on 28nm at TSMC. The really interesting question is what process Carrizo is going to be built on?
CPU efficiency and the foundry question
Right now, all of AMD’s “big-core” x86 CPUs are built at GlobalFoundries. GF isn’t rolling out multiple types of 20nm this time around, but is pursuing a unified strategy of offering one type of silicon that can stretch to cover multiple targets. This 20nm LPM node (shown below) may indeed be able to hit a wide range of targets – but it’s unlikely to stretch far enough to accommodate the full range of AMD’s product line. That’s probably part of why we see TDPs yanked down to 65W for Excavator.
Intel can hit wide TDP ranges on its process nodes partly because it enforces extremely strict design controls on the engineering team. GF, being a foundry, can’t exert the same amount of control over its partners, which means AMD has to find ways to fit within the 20nm process if it wants to build cores at GloFo. But that throws the question back to the CPU team – can they design a Bulldozer-derived part efficient enough to fit within the power envelope?
Up until now, AMD’s big-core business has relied on pushing TDPs higher to hit better performance targets. The company’s six-core and eight-core products all draw more than 65W. Moving to a 20nm low-power process at 65W means Excavator’s performance-per-watt needs to leap forwards relative to Kaveri. This, in turn, may partly explain why AMD is keeping the GPU core as a GCN-derivative – it may make sense to retain that architecture as a known quantity while negotiating a die shrink and architectural shift.
All of this assumes that AMD is keeping its upper-end x86 business at GlobalFoundries, mostly because there’s no reason to suspect that business could shift elsewhere. Still, I think it’s clear that AMD is strategically re-evaluating the product segments it participates in. We don’t see a 20nm version of Kabini, Beema, or Mullins on this roadmap because AMD is ramping up an ARM core instead. If that core does well, it may become the poster child for AMD’s new push into different market segments. The 14nm and 20nm tape-outs that Lisa Su discussed at the company’s last earnings call could refer to designs for the PS4 and Xbox, not merely the PC side of the business.
AMD can’t afford to ditch the x86 business altogether at this point, but we know it has huge plans for embedded and semi-custom, which Rory Read has talked about as accounting for up to 50 per cent of the company’s revenue in future years. This roadmap reflects some of those plans – with limited cash to spend, AMD is going to put more funds into emerging markets, shore up its overall x86 position, and push forward with second-generation Kabini hardware while keeping the entire line-up standardised on the GCN architecture.