Discussing the future of AMD has almost become a global sport, with everyone chiming in with their ideas of how things will play out or how they should be run. I’m guilty of that at times as much as anyone else – ideally I’d like to see AMD succeed and return to the days when the Phenom II and K10 and K10.5 architectures were still competitive with the Core 2 and Core families from Intel, but it isn’t a realistic possibility. The landscape has changed so drastically that those golden years of the late 90’s and 2006 specifically (when Intel couldn’t make enough Core 2 Duo processors to satisfy demand) are long gone. Intel had their struggling time with the Pentium 4 and they walked out of that one relatively intact.
It’s AMD’s turn now to close the book on Bulldozer and all its derivatives and look to the future, to a fresh start.
Moving to a new socket and DDR4
One of the most anticipated changes in the history of the computer is the move to DDR4 memory, primarily a change spurred on almost solely by Intel. Intel’s Haswell-E HEDT line is probably the only product on the market where the switch to DDR4 could go smoothly. It’s a low-volume segment with high profit mark-ups to make the effort to manufacture products for enthusiasts a worthwhile endeavour. It’s to AMD’s benefit that they let Intel tackle this on their own, because they can step back and learn from what comes out of the launch and the issues that early adopters face. In the meantime, work continues on their next generation of products.
Moving to DDR4 for AMD necessitates replacing not two, but four sockets. They can’t immediately do this for all of their products because it would be a financial headache, not to mention a strain on the limited resources they have. So what’s likely going to happen is they will be moving their APU line on the FM socket to DDR4 first. Socket AM1 already benefits from APUs with super-low TDPs, so bringing DDR4 to that platform doesn’t change things much. Nothing can save AM3+, so it’s better to just let it stagnate and die off on its own while they work on a replacement.
AMD’s Opteron server line would certainly benefit a lot from the memory switch as well, but they are currently working on Project Skybridge, which merges the SeaMicro Fabric technology with a socket that, for the first time, offers pin-compatibility with ARM processors that AMD has licensed out. AMD’s been doing a lot of work to make their GPUs and chipset as agnostic as possible to allow them to fit in whatever processor cores fit the intended use better. It’s not going to come immediately for that platform, though – validating new tech for use in servers and other enterprise applications typically takes years. We’ll only see more of Project Skybridge in mid-2015, with some samples already being sent out to select companies that would be partnering with AMD for their Seattle ARM server family, which already uses DDR3 and DDR4 memory controllers.
So while Intel will hold the keys to DDR4 for this and next year thanks to Haswell-E and Skylake, AMD will play their cards closer to the chest, popping out their answer in 2016. Waiting an extra year won’t damage their cause much, because DDR4 prices will be sky-high for a while still.
The fresh start we’ve been waiting for
When Jim Keller moved back to AMD in 2012, many people including myself were interested in what the big plan was. Keller was on the team that created the first Athlon FX processors and he’s also one of the principal designers of the x86-64 standard. In the video above, Keller talks about the results of the work being done with Skybridge and the lesson’s they’re taking away from working with ARM architectures and putting into the final product. AMD has proven before with Bobcat and the new low-power Jaguar processors that appear in the Xbox One and Playstation 4 that they’re easily capable of building something that is competitive with Intel and, to some extent, ARM.
At around 9:50 there’s a little more detail into how they approached Skybridge and the pin compatibility for the APUs. Keller’s team took the Cortex-A57 architecture, compared it to their new x86 designs and made them physically similar to each other – similar enough that switching between ARM and x86 only requires a single component change, making production cheaper for AMD and leaving space open for updating the Skybridge platform for newer architectures.
This is a calculated play by AMD to make sure they’re ready for whatever the future brings. Lets say that Android invades the laptop and embedded space more thoroughly and begins to take away tiny amounts of market share from AMD and Intel’s traditional products – AMD goes on to sell their new APUs to these OEMs and maintains or grows market share because all that’s now required is to use a new version of Android that supports these APUs and the rest of the chipset.
No matter which processor core the customer chooses, they still get to use AMD’s GPUs for GPU compute operations and they gain access to embedded technology like gigabit Ethernet and AMD’s latest creation, the TrueAudio DSP.
If AMD hadn’t sold off Imageon to Qualcomm back in 2009 they wouldn’t have to do this today, which is basically all you need to tell you about how ex-CEO Dirk Meyer approached the company’s future plans at the time. Meyer didn’t think that mobile would grow as big as it has done today and he certainly didn’t anticipate an ARM invasion into the server market. He was the last push to the Bulldozer architecture and the current mess AMD is in today, having never quite completed the financial recovery he was looking for from the high-performance desktop market.
No-one likes Bulldozer’s weak single-thread performance, no-one likes that it requires lots of power and produces so much heat and certainly few people are fans of the module design with shared hardware between integer cores. Wiping the slate clean gives Keller a chance to revisit old ideas that worked as well as bring in fresh, new ones from his experiences in building chips at Apple.
On-die memory doesn’t really fit into their plans
There have been several reports on the internet from the same source, Bits n Chips, which says that AMD is thinking of using memory chips much like the eSRAM on the Xbox One’s APU. Intel’s currently doing the same for their Core i5-4670R and i7-4770R processors with Iris Pro graphics, with both chips stacking 128MB of eDRAM into the same die for super-fast memory access. There is some benefit to this because you don’t have to change around your memory controller to improve performance, but it’s definitely something that needs to be specifically targeted by developers.
That memory could be used for graphics or even as a general cache, but it introduces an extra layer of complexity for AMD that they might not want to deal with for a desktop consumer part. Making the Xbox One APU is a little easier because they only have to make a single SKU, but when you have to mass-produce chips that need to filter into different price points you don’t want to be making something that needs to sell in huge volumes (larger than what you’re used to shipping) and be unable to sell it effectively.
If they’re doing this and not going into DDR4 memory as soon as possible it would still also require a socket change. Why keep more or less the status quo for a product that is starved and will still be starved of raw memory bandwidth? Sony’s Playstation 4 shows that a single, large, uncomplicated pool of GDDR5 memory is easier for developers to take advantage of than the combination of slow DDR3 and eSRAM.
I’m pretty sure that AMD knew all of this well in advance of both console launches, so they have access to the data that would tell them definitively, with a pretty large sample size, how difficult making these things would be if they had to sell a similar design to the general public. On-chip memory isn’t a very likely play from AMD because they need to make APUs less complex to manufacture, not more.
So you’re stuck for an ugrade, bra?
So let’s say you’re on socket AM3+, like me, what do you do? Well if you’re on a Athlon X2, X3 or X4 processor, upgrading to something like the FX-4300 or FX-6300 will still give you a nice jump in CPU performance and you’ll be benefiting from the L3 cache. If you’re on a Phenom II processor, you’re pretty much better off just sticking with what you have. The K10 architecture may be older, but it still can perform pretty closely to what AMD’s offering from the Piledriver family.
But don’t expect anything new, ever. AM3+ is dead weight and needed to be killed off. Even the chipset has ties to hardware that was released a decade ago. The same goes for socket FM2, which is a lot easier to squeeze out of because you can keep your existing FM2-compatible APU and move to a FM2+ motherboard. You gain a more modern chipset and features if you move to the A88X platform plus you have access to Kaveri and eventually Carrizo APUs which will be based on the Excavator architecture.
If you’re on socket FM1, you already know that it’s dead. If you’re on AM3, you’ll know by now that it’s been dead for years. Instead, it’s far more ideal to jump to the socket FM2+ platform in this case or begin looking at Intel’s Haswell parts to give yourself a nice boost in performance and efficiency.
If you’re on socket AM1, good news! You can stay right where you are. AMD has plans to support it for the next two to three years and because the APU is a full SoC (System on Chip) all that’s required to take advantage of newer technologies is a chip upgrade. APUs based on Jaguar refresh will be likely released soon after CES 2015 and interesting to toy with. By the time 2016 rolls around it would be cheap enough to make a solid case for an upgrade to AM1+ and DDR4 memory.
So tying this all up…
I have a family friend who regularly asks me what’s happening in the tech world when he visits. Without fail, every time he mentions something that he read in a magazine or online about AMD dying or being dead or abandoning the desktop market and to be honest, it’s a function of the fear-mongering that some tech journalists are doing to reduce confidence in AMD’s ability to stay in the market. Let’s face it – when the console market embraces your products almost completely and when you have one of the most flexible low-end products we’ve ever seen (Intel’s Bay Trail notwithstanding because the chips are soldered in) there’s little chance that you’ll disappear overnight.
Rory Read has plugged most of the bigger holes that were draining money, they’re doing as much as they can to improve developer relations, Lisa Su is awesome, the server market is interested in a better-supported ARM architecture and there is still enthusiasm over OpenCL GPU acceleration, even if that market is only slowly building up a following. There are some very smart people now at the helm of the company and they are steering it in a direction that I like, even if their products are marked up ludicrously in our local market.
I still cannot justify putting down over R2700 for the A10-7850K when that money plus R50 extra buys you a Intel Core i3-4150 (so we’re already up on CPU performance) plus a Radeon R7 250X with 1GB of GDDR5 memory (massively improved GPU performance).
It will be a while before they get back on their feet again. AMD doesn’t have a process advantage over Intel, but that could change in the future thanks to new technologies finding their way to Global Foundries and perhaps the use of TSMC’s 20-nanometer production process. They clearly have the right kind of engineers to design a CPU architecture from scratch and they’ve learned a bunch of positive lessons from Bulldozer. They’ve been able to almost completely secure the console market and they have a foothold in Apple’s iMac Pro because Nvidia won’t sell Apple a custom GPU for use in the swanky new chassis.
All they need to do now is figure out how to bring out their own “Core 2″ processor, a chip that sets them up for better success in the CPU market and doesn’t give Intel fans a hernia when they begin arguing over whether a dual-module chip can be considered a quad-core processor. Don’t believe all the negativity and doom surrounding the company on the internet, AMD’s doing just fine now. They’re down, but not out for the count.