How AMD plans to get back in the game

AMD admits it has "lost some share" in computing and graphics. But it says new technology, a simpler product roadmap and more focused strategy will make it competitive once again in high-performance desktops and servers
Written by John Morris, Contributor

AMD has spent the last few years cultivating new markets for its CPU and graphics technology. It has had some successes, most notably the Xbox One and Sony PlayStation 4, which both use its processors. By the end of last year, these "growth markets" in the enterprise, embedded and semi-custom represented 40 percent of the company's revenues.

But part of the reason is that AMD's core computing and graphics business has not been growing. The company's share of the server market has dwindled to a few percentage points and its heavy exposure to the consumer market left it especially vulnerable to the PC downturn. More recently, rival Nvidia has been stealing share in graphics with its highly-efficient Maxwell GPUs.

This week, at the first analyst day AMD has held in several years, the company explained how it plans to get back on track. AMD announced several new products, rejiggered its roadmap, and vowed to deliver more competitive technology to regain share in high-performance PCs and servers.

"We have lost some share, I recognize that," CEO Lisa Su said. "We will gain that share back in more profitable segments."

AMD traces many of its current issues back to the Bulldozer architecture that debuted in late 2011 and its big bet in coming years is a replacement code-named Zen. The company still isn't providing many details on Zen, but it did state that the CPU core is designed specifically for higher performance, supports multi-threading and will deliver 40 percent more instructions per clock. The jump from the 32nm and 28nm nodes with planar transistors to a FinFET process, presumably GlobalFoundries' 14nm technology, should also deliver a boost in performance and power-efficiency.

It's good to see AMD committing to more competitive CPU performance. But it's worth noting that AMD isn't saying it will leapfrog Intel; it is merely saying that Zen will narrow the gap, or as CTO Mark Papermaster put it, bring them back in the "zone of competitiveness."

The other issue is that the rollout of Zen seems to be slower than expected. Surprisingly it will appear first, sometime next year, on the FX Series desktop CPUs--a product line that AMD has largely neglected for the past couple of years to focus on mobile APUs--for PC enthusiasts and gamers. This will be followed at some point by desktop APUs and then mobile APUs.


During the event AMD announced its latest 7000 Series low-power "Carrizo-L" APUs, which are based on the older Puma CPU design. Later this year AMD will announce its sixth-generation A-Series mobile APU family, known as Carrizo, which is already in production and will be in laptops in time for back to school. This one uses a newer Excavator core, which provides incremental improvements in power efficiency, "but that wasn't going to get us back to where we needed to be on performance," Papermaster said. The desktop APU line, known as Kaveri, remains unchanged this year though AMD cut the prices yesterday.

AMD is planning to release seventh-generation mobile and desktop APUs in 2016, but for now it won't say whether or not these will use FinFETs or the new microarchitecture, so it is possible we won't see Zen in mainstream desktops and laptops until 2017. By that time Intel will likely have moved on to 10nm processors.


AMD is also counting on Zen to "re-enter" the server market, as executive Forrest Norrod said. The next-generation Opteron will use "many" Zen CPU cores and deliver "disruptive" memory and I/O bandwidth. It is listed on the roadmap for 2016-2017, but at one point AMD mentioned it would be "sampling" next year, so it is likely to be a 2017 product. As rumored, AMD is also developing a server APU with graphics that will deliver multiple teraflops of performance for high-performance computing and workstations.

The micro-server market has not developed as quickly as ARM or customers like AMD anticipated. Last year, AMD released an early version of its Opteron A1100 (Seattle), with four or eight Cortex-A57 cores, for developers. The processor will ship in the second half of this year and will be used in storage products, Internet of Things gateways and Web front-end servers.

AMD had previously planned to follow this with a family of 20nm ARM and x86 processors, including a low-power A57 SoC, that use a common platform (Project Skybridge) in 2015. But CEO Lisa Su said that the 20nm process fell short of expectations and customers saw little need for single platform for both x86 and ARM processors. AMD's custom ARM core, code-named K12, has also been pushed out to 2017.

AMD also announced new Radeon M300 mobile GPUs, which are available in laptops from Alienware, Lenovo and Toshiba, as well as new Radeon 300 Series desktop GPUs for OEMs starting with HP. These are based on existing Graphics Core Next GPUs, not new technology.

The more interesting announcement comes later this quarter when AMD launches its new high-end Radeon R9 300 Series GPUs. These will be the first desktop boards to use High Bandwidth Memory (HBM), stacks of DRAM memory placed next to the GPU in the same package, which AMD says deliver three times the performance per watt and use half of the power of separate graphics memory (GDDR5). The initial HBM devices will stack four layers of DRAM to reach capacities of 1- or 2GB, followed by eight-layer versions with twice the density. So a GPU surrounded by four HBM stacks could have 8- to 16GB.


Next year things will get even more interesting when AMD rolls out a next-generation GPU, manufactured on a more advanced process with FinFETs and using second-generation memory stacks. HBM2 will have twice the density and bandwidth, and overall this GPU will deliver twice the performance per watt, AMD said. However, Nvidia's Pascal GPU, due in 2016, is also expected to use HBM2 to pack up to 32GB of high-bandwidth DDR DRAM in the package. The combination of FinFET transistors and a new high-bandwidth memory architecture should deliver a significant leap in GPU performance for both 3D graphics and general-purpose computing in 2016.

For AMD it comes down to execution. The company spends a fraction of what Intel does on research and development--and less than even Nvidia does--so it has been forced to simplify its roadmap and delay or drop several products. The Radeon R9 390 should get AMD back in the game, but graphics will continue to be a horse race. More important, if AMD can deliver a more competitive CPU, it should win back some business in desktops next year and in servers in 2017.

Editorial standards