Enthusiasts barking up the wrong tree
Before diving into Read's remarks, I think it's important to address the fundamental difference between how a lot of AMD's historic fans see the company versus the direction AMD is now charting. The enthusiast community has a long tradition of viewing AMD's performance through Intel and Nvidia-specific lenses. There's nothing wrong with this, as such, but it misses a critical fact: when Rory Read took over as CEO of AMD, that narrow focus, combined with multiple execution problems, had nearly killed the company.
AMD still cares about its traditional markets, but they aren't driving the entire structure anymore. Going forward, AMD believes embedded computing, dense servers, high-performance GPUs, and professional graphics are just as important to its profitability. Read's goal is to move AMD away from a model where the company relied upon lightning in a bottle strategy and only turned a profit two years out of eight.
Products and process
In the past it was assumed that AMD would automatically move to new process nodes as soon as those nodes became available. One thing Read emphasized was that 28nm would remain a major node for AMD for quite some time to come, saying:I don’t think we’re at the peak of volume for 28-nanometer. We’re going to continue to leverage that and I think there is great opportunity to make money on that... The mix for us has to be where we can get the yields and the business like game consoles, semi-custom, embedded, and even in client, 28-nanometer is going to be just fine for the next several years.
In his very next sentence, Read seems to reverse himself, saying "Now, will we move down to Fins? Yes, absolutely, and our next generation products go there and as we introduce them in [20]16."
Here's what this means, in aggregate: AMD is planning to invest in next-generation process nodes at 20nm and 16/14nm FinFET, but only on certain products and only when the company believes it can recoup substantial investments from doing so. AMD's presentations to date have hinted that when its combined ARM/x86 platform debuts next year, both Jaguar and its Cortex-A57 chips will be built on 20nm, incorporate HSA support, and use a common graphics core. This implies that AMD's low-power x86 chips could see a nice uptick in performance year on year.
Similarly, Read hinted that K12 and AMD's upcoming Zen platform will move to 16nm and FinFETs fairly aggressively (if they don't deploy on those technologies to start with). Big cores like the current Kaveri, FX-series, and upcoming Carrizo, on the other hand, well -- those are precisely the volume parts that are going to stay back on older process nodes.
We actually see signs of this strategy already. Trot over to NewEgg, and the 17 APUs on sale there aren't all Kaveri or even all Richland -- AMD still has a healthy number of Trinity-class parts on the market. In the Dirk Meier era, AMD's strategy was to shift to the entire stack to a new node as quickly as possible in order to maximize cost savings on a new node. What Read has laid out is a different plan, in which AMD only moves to new nodes when it makes sense to do so for each particular class of parts.Die stacking, tablets, and the 2015 timeline
One of the most consistent rumors flying around AMD is that the company would deploy die stacking technology with Tonga or possibly on Carrizo. AMD's HBM (High Bandwidth Memory) is under significant development, but the only thing Read would say at this juncture was "There will be some very interesting technology in terms of stacking and [chiclets] (ph) to kind of -- you could go cutting edge on certain parts of the design and then combine it with more efficient designs to really get a better cost model." The key here, however, is that he said this explicitly as part of a discussion on cost optimization to avoid needless expense.
AMD is working with Hynix on High Bandwidth Memory, but don't expect a near-term debut
When asked if he'd like to take AMD into tablets, Read's response was that while this was obviously an important long-term market, Intel's aggressive marketing practices (referred to as wrapping a $20 bills around every one of their processors) made it impossible for AMD to aggressively chase share in that space without incurring crippling losses.
Where do all these plans leave us for the back half of 2014 and into 2015? The hard truth: mostly in a holding pattern. The back half of this year and early next year will be GPU-centric, with multiple new Mantle games launching and a possible new GPU family in the next six months.
Next year, the big story will be Skybridge, dense ARM servers, and whatever more modest benefits the updated Carrizo platform can offer. It's possible that AMD will introduce its first iteration of stacked memory (HBM) in 2015, but it wouldn't surprise me if that update slips into the 2016 timeframe. Expect it to launch on GPUs first -- the challenge of moving to a combined ARM-x86 platform and bringing up two new CPUs is a sufficiently large hurdle to tackle without simultaneously taking on HBM on the APU as well.
Nothing was said about further updates to the AMD FX platform, suggesting that even if AMD keeps that brand name around when its Zen architecture debuts it won't be an update to the current AM3 platform. After Rory Read's comments on optimizing value at particular nodes we don't expect to ever see a Bulldozer-derived processor launch below 28nm.
A different AMD
I wasn't certain what to make of Rory Read when he first took over at 1 AMD Place. His engineering background was clearly lacking and his answers to technical questions were often painful to read. Three years into his tenure, however, I think Read has proven himself -- the AMD of 2014 is in a far better overall position than the AMD of 2011.
Part of saving AMD, however, required recognizing that it could no longer depend on its traditional markets to sustain it. The branch into dense servers, ARM architectures, semi-custom embedded hardware and a newfound focus on the professional GPU market means that AMD wants to compete in spaces beyond those where Intel and Nvidia have been its historic competition.
As an enthusiast, it's frustrating to hear that AMD won't pursue node shifts as aggressively and plans to optimize its business for cost and volume with strategic implementation of cutting-edge technology. The bottom line, though, is that the company's previous practices were flatly unsustainable. Attempting to fight solely against Intel and Nvidia didn't give AMD a focused advantage, it left the company exposed to every bruising market downturn or shift in competitive performance.
I don't think AMD has permanently walked away from trying to compete with Intel, but Read and his executive team correctly realized that the company needed more income diversity if it was ever going to tackle Chipzilla successfully. The downside to this strategy is that classic AMD enthusiasts who have been hoping for some major breakthrough on the high-powered APU front will likely be waiting into 2016.
