The roadmap for upcoming AMD products has been pretty murky lately, but the company revealed some of its key plans for the next few years at its Financial Analyst Day today.
The zen of Zen
The biggest news has to do with Zen, the new x86-compatible CPU core meant to replace the current Bulldozer family. Today's presentations didn't go deep into the technical details of the new architecture, but what we learned was largely encouraging. We know the general timeline and outlines of the first Zen-based products.
First of all, Zen is indeed a new high-performance core intended to compete against Intel's best x86 processors. AMD expects this core to deliver about 40% higher performance per clock cycle than today's Bulldozer variants, as the slide above indicates. Zen looks to be more of a "brainiac" architecture like K8 and Broadwell than a "speed demon" like the Pentium 4 and members of the Bulldozer lineage.
Also, the first Zen-based cores are just the beginning. Future "Zen+" variants of this architecture should improve per-clock performance even further.
The Zen core will feature simultaneous multithreading (SMT), or the ability to track and execute multiple threads per core. Although SMT can extend beyond this limit, Zen's version of SMT will stop at two threads per core, like today's big Intel cores. The inclusion of dual threads per core follows a proven template for success in big x86 CPUs, and it also should put AMD on more equal footing with Intel from a marketing standpoint.
The other reassuring technical detail is Zen's addition of a "high-bandwidth, low-latency cache system." Bulldozer's caches have been considered something of a pain point since the first chips arrived, and they clearly don't perform as well in directed subsystem tests as the caches on recent Intel CPUs. Thus, this little nugget also seems like good news.
AMD expects to bring the first Zen-based silicon to market in 2016, and those chips will be based on a chip fabrication process that uses FinFETs, also known as 3D transistors. The use of FinFETs is positive news in the sense that it should allow for faster switching speed and lower voltage operation than traditional planar transitors—and AMD's major competitor will be producing chips on its second or third generation of 3D transistors by 2016. AMD didn't reveal the specific foundry or process on which these chips will be made, but the obvious list of candidates is pretty small, including the 14- and 16-nm FinFET processes at Samsung, GlobalFoundries, and TSMC.
A many-core processor for high-end desktops and servers in 2016
What's really intriguing about the first Zen-based silicon is the sort of systems it will target. My sense is that AMD will produce a single high-end CPU and aim it at two markets: high-end desktop PCs and servers. The roadmaps for Opteron and FX processors describe a chip based on the Zen core with a "high core count with multi-threading," ample memory bandwidth, and lots of built-in I/O.
On the desktop side of things, the Zen-based FX processors will be supported by a new AM4 platform that includes support for DDR4 memory. This same AM4 platform will also support smaller APU chips, unifying the company's desktop offerings around a single socket.
Into 2017 and beyond
The downside of the news about a big, Zen-based chip in 2016 is that AMD apparently had to prioritize this one chip over other options. The company's 2016 APU products for desktops and mobile systems will not yet incorporate the Zen core. Also, the schedule for the K12 core, the ARM-compatible sister to Zen, has been pushed back to 2017. CEO Lisu Su further revealed that Project Skybridge, the effort to make ARM- and x86-compatible CPUs share the same sockets and motherboards, has been nixed, reputedly due to lack of customer demand.
What we know of AMD's plans for 2017 looks intriguing. One chip in particular, mentioned in the datacenter roadmap, appears to be something different from anything we've seen before: a "high-performance server APU." This product may be the first from AMD to combine a best-of-breed x86 CPU with truly powerful graphics on the same chip. The major limitation preventing this sort of converged product in the past has been the bandwidth available in a CPU socket. Simply put, CPU sockets haven't had the capacity to support a data-streaming processor like a GPU.
The CPU-socket bottleneck may be alleviated by the "transformational memory architecture" mentioned on the roadmap, which is almost certainly the high-bandwidth memory (HBM) subsystem that AMD has been developing for use with its GPUs. HBM situates the memory on the package next to the chip it serves, such as a GPU or an SoC, and promises substantially higher bandwidth than traditional DRAM in combination with lower power consumption. The total physical size of the chip-plus-memory solution can be much smaller with HBM, as well. AMD's execs strongly hinted that HBM will make its way across the company's products going forward, and this server-class APU looks to be one beneficiary.
That prospect is exciting because it could give AMD a foothold in markets like HPC and supercomputing, where it hasn't really mounted a credible challenge to rivals like Intel and Nvidia in recent years. This big server-class APU could also begin to deliver on the promise of AMD's HSA programming model, which has so far been full of promise but lacking truly compelling hardware implementations.
This big APU won't just target servers, either. AMD hasn't officially extended its desktop roadmap into 2017, but a source familiar with the firm's plans has indicated to us that a beefy APU of this sort could make its way into client systems like laptops and desktops, as well. The vision here is for a single-package product with HBM to combine a high-performance CPU with a Radeon graphics solution that has real gaming chops—something well beyond the offerings we've seen from AMD's own APU lineup and Intel's Core i5/i7 processors to date. If it works out well, such a product could give AMD a product unlike anything else available: a decent gaming PC in a compact package, ready for gaming laptops, all-in-ones, and small-form-factor systems.