AMD is going to be in a very tight spot in the next two years, coming up against the launch of Intel’s Broadwell, Skylake and Haswell-E chips that further extend big blue’s lead over their competitors. With AMD concentrating on their low-power APUs, improving ARM architecture and getting Excavator on track for a release in 2015, there doesn’t seem to be a lot of time for the company to come up with new, high-performance offerings to replace the aging socket AM3+ and the FX processor family and there definitely doesn’t seem to be any indication that AM3+ will see any more action either. 2016 is their year to get things back on track and it’s beginning to look very interesting already.
2016, however, is a long way away. Whatever AMD’s planned for the desktop and performance markets in the meantime is completely different from what’s presented here today. Until then, they’ll have to fend off the advancements by Intel into BGA-only processors, DDR4 memory and whatever graphics improvements they make to the Intel HD family. While the plans to meet those problems are being worked on or in the process of being acted on, AMD’s chosen to reveal some of their cards early.
The first card was a quick look at what kind of intellectual property (IP) AMD is working on. They have three focuses at this point – the Crown Jewel cores, the Complementary System-on-Chip (SoC) functions and the optimised software produced with the aid of their software partners. It’s a pretty wide net to throw out and the plan hinges on all three aspects resolving themselves properly. At the core of their plan (AMD has this obsession with planets now!) is the CPU, GPU and ARM-based cores that will be doing most of the hard work. This signifies that much of AMD’s work will be going into integrating technologies into their hardware instead of relying on software workarounds to issues that they’ve identified.
Moving up is the SoC IP, another part of AMD’s focus. This is a big departure from what they do now, which is leave most of the functions found on a SoC design to the chipset, also known as the South Bridge. AMD, then, is moving into a product that is more fully packaged and modular, but at the same time facilitates upgrades and performance increases simply through swapping out the APU – at least this is what the plan here looks like to me.
Finally, AMD’s going to use the foundation laid by these two plans to build up their optimised software stable. This includes improvements to Windows and Linux operating systems, a bigger focus on Android and then updates and improvements to APIs like DirectX, OpenGL and CL as well as the in-house Mantle API. How AMD leverages all three levels to make products for different markets will differ widly, but they’re all going to use the same core technology.
One of the big announcements for the future is Project Skybridge, an ambidextrous chip design debuting in 2015. Skybridge is interesting for two reasons; one is that it will be the first commercially-available APU made by AMD that has a big focus on mixing in ARM and x86 architecture into a single package. AMD says this is a fully HSA-compliant design, which means that such a chip with a mixture of ARM and x86 cores could be programmed to perform functions on the separate Compute Units (CUs) based on what hardware the code performs better on.
The other reason is that the two technologies will be pin-compatible. In electrical terms, this means that there’s a large amount of leeway in terms of how a Skybridge chip is built up. It could be using all of the complimentary SoC IP and hardware with just ARM CUs, or it could be heavily sided to ARM with one or two x86 CUs for running code that doesn’t run well on the ARM chips. Equally, AMD could offer Skybridge APUs with a focus on x86 performance, but with one or two ARM CUs for running background code and security software separate from the main OS.
While that’s pretty awesome on its own, both ARM and x86 will also be paired with Graphics Core Next (GCN) CUs. What this means is that for developers making a HSA-aware application, they can rest easy that no matter what AMD APU they target, it’ll always include at least two types of CU hardware that they can target for hardware acceleration of HSA software. That’s a big boost to adoption of HSA-compliant software in the future because the chicken-and-egg approach has always requires the hardware to be present at all times.
Its also a big change for the mobile, notebook, desktop and server segments. No matter which platform you target, they’ll all have some HSA-compliant hardware. With things as they are currently thanks to the HSA foundation, there’s not too much support for HSA-driven software on the tablet and mobile phone market. With Skybridge providing the foundation for this that isn’t currently offered by any other vendor (Intel included) AMD could be eking out a very solid spot for themselves in all four markets.
The roadmap for the future, then, is pretty much covered in ARM developments. Products in 2014 and 2015 feature Cortex A57 processors, but these will be minorly customised designs that work off ARM’s own designs. The big kahuna pops up 2016 and it’ll be a monumental feat of AMD can pull it off – two new architecture launches, one on the x86 side and the other is a custom ARM architecture designed completely in-house by AMD.
Effectively, this means that we’re going to be seeing Excavator cores for socket FM2+ in 2015 with 2016 heralding the end of the Bulldozer (and all of its derivatives) architecture. AMD needs to make the new ARM-based K12 architecture pin-compatible with their x86 designs, which isn’t currently possible with Steamroller or Piledriver. What will come out of it? Probably something very similar to the Jaguar and Puma architectures, both of which have much potential for further scaling into higher TDPs and different form factors.
That also puts some interesting things into perspective – in 2016, most mobile devices need to support some form of connected standby to function properly with whatever form of Windows and Windows Phone will be out then. If AMD can make their architectures compatible electrically, they also have the opportunity to work on a x86-ARM translation layer. Its entirely possible to have the ARm cores take over when the device enters into connected standby or an idle mode, dropping power usage really low. When more power is needed, the x86 cores are woken from their slumber and take over the reins to allow the ARM cores to work on their own thing or switch off.
2016 will be an interesting year for AMD and its fans and I look forward to seeing what they’ll be able to achieve thanks to HSA. What are your thoughts about Project Skybridge and the greater concentration on ARM developments? Let us know in the comments below!