AMD's desktop and mobile roadmaps
Under Rory Read, AMD's desktop CPU roadmap languished. The only meaningful improvement to high-end CPUs under his leadership was the FX-9590 -- a 5GHz Piledriver-class core with a 220W TDP. In 2016, that's finally going to change, thanks to the launch of AMD's Zen. AMD hasn't given a date on when to expect Zen (scuttlebutt has pointed to second half of the year, but AMD hasn't confirmed or denied that estimate).
We've expected that Zen would come to desktops, but CEO Lisa Su confirmed that Zen would actually ship on the desktop first. The CEO noted that this market has proven to be important to AMD's overall market position and design wins, and AMD is committed to attacking this space.
It's assumed that Zen will also power the 7th-generation of desktop and mobile APUs that are shown here, though AMD only explicitly stated that Zen was coming to the desktop. One major change coming in 2016 is that the APU and CPU will occupy a common desktop socket, AM4. In the future, you won't have to choose between two different motherboard sockets when you pick either an APU or a CPU from AMD -- you can move from one to the other.
If AMD can deliver the estimated 40% IPC uplift that it's claiming, it should be far more capable of challenging Intel in every segment of the market. Even if Zen isn't capable of going head-to-head with Core i7 (and we aren't saying that it won't), a 40% IPC jump would still give AMD a much more competitive position against the Core i5 and Core i3. The days of needing two or more AMD cores to match a single Intel core with Hyper-Threading enabled could finally be drawing to a close.
One point AMD made indirectly is that its FX chips will be CPU-only, while the APUs will continue to include a graphics core. With a common platform, customers should have more freedom to shift to solutions that meet their needs. AMD did confirm, in its subsequent Q&A session, that the era of the "cat" cores is over. The Jaguar / Puma family of chips doesn't make the jump to FinFET.
AMD didn't say much about graphics directly. Fiji is due to ship out in the second quarter, but the company didn't reveal anything about the GPU core inside the 7th Generation APUs that are set to debut next year. We know that Carrizo's GPU is based on Tonga and we've heard rumors of a Greenland GPU core that would finally overhaul the GCN architecture. We do know that AMD expects to adopt FinFETs for its graphics solutions next year, which means the company will move to either 16nm or 14nm at TSMC or Samsung respectively. Second-generation HBM is also set to debut, which could deliver up to 8192-bit memory interfaces and a further bandwidth increase.
Also, Lisa Su noted that while AMD had started multiple 20nm designs, those projects aren't going to come to market, due to minimal profitability on that silicon. Back in 2012, we published slides from Nvidia that collectively illustrated that 20nm wasn't a useful node for high-performance silicon. While we saw some mobile uptake of that node, it's worth noting that we were absolutely right -- if you're in the high performance computing business, TSMC's 20nm was, in fact, "essentially worthless."AMD's server roadmap
AMD's first ARM-based server, the A1100 Opteron, will debut in the second half of 2014 2015. This delay, according to AMD, was driven by the market -- ARM microservers simply didn't explode the way AMD thought they would initially, and it decided to delay its own efforts as a result.
The K12 is expected to debut in test silicon next year as a follow-up to the Cortex-A57 based Seattle, with an emphasis on high performance at the upper end of the ARM computing market. Exactly how it will line up against future Zen-based Opterons, and the degree of overlap we may or may not see between them, is still unclear. Mainstream shipments aren't expected until 2017, which likely gives the ARM server market time to mature. AMD used to talk about dense servers accounting for up to 25% of the server market by 2017, but the company tacitly acknowledged today that these estimates were a tad optimistic. The failure of that market to materialize is likely one reason AMD junked its SeaMicro acquisition.
The most interesting statement in AMD's server presentation is the "Disruptive Memory Bandwidth" and "Transformational Memory Architecture" claims. We've seen rumors before that AMD might integrate HBM into APUs. If the company were to do this, it might make the most sense to integrate it into servers first. While HBM has been discussed as a game-changer for integrated graphics -- and it truly could change the rules of the game in that segment -- it's important to stair-step new technologies into markets that can afford the additional cost.
Offering HBM on a server chip would give AMD access to a nearly on-die cache that would offer vastly improved bandwidth compared with traditional DRAM. AMD likely can't afford to take Intel's route of building a 128MB L4 cache on die, but an HBM memory segment (backed by a conventional DDR4 main memory) could be a potent alternative.Putting it all together:
Not much of what AMD revealed today is going to change forecasts for 2015. Carrizo is expected to offer better battery life over older Kaveri systems, and the R9 3xx updates should put AMD on much better footing against Maxwell. But the bottom line is that for many of AMD's businesses, 2015 is a bit of a pause. Even in situations where we see new product launches, like the upcoming Seattle ARM core, these are the vanguard launches of product segments where further products, like K12, are expected to make a larger difference.
It's clear that 2016 will be the make-or-break year for AMD. The entire future of the company is riding on two premises: One, that it can build a business around K12 and that investing in a custom ARM architecture will pay off over and above the price of continuing to license a standard ARM core. Two, that its upcoming Zen CPU will offer vastly improved performance per watt compared with the Bulldozer-class chips, scale across the entire industry from mobile to server, and offer a compelling alternative to Intel's own chips in each of these segments.
It's not hard to believe that Zen will beat out Bulldozer in terms of instructions-per-clock: AMD could've beaten Bulldozer's IPC by shrinking Phenom II to 14nm and adding a few BD-derived features to the core. Hopefully AMD's confidence that it can offer a top-to-bottom solution against Intel is well-deserved.
