The launch earlier this year of the first Zen processors was a big step to make AMD relevant again. It demonstrated that the company could produce a competitive CPU design manufactured on an advanced process. But AMD knows it needs to do more to win back business from Intel and Nvidia.
At its analyst day, the company announced more new products and a roadmap to remain competitive for years to come.
The most significant news is the confirmation that the Zen server chips code-named Naples will ship in June under the new Epyc brand (RIP, Opteron). Epyc has up to 32 cores (64 threads) and is designed for mainstream single- and dual-socket servers -- the bulk of the market and an area that AMD has largely surrendered in recent years because it lacked a competitive product.
Each processor has 8 memory channels and 128 lanes of PCI-Express 3.0; a two-socket server with two Epyc processors connected using its Infinity Fabric can hold up to 32 DIMMs or 4TB of memory. The first images of the Epyc showed how AMD managed to produce a 14nm 32-core processor by slicing it up into four chips with higher yields packaged into a single chip -- a direction that Intel has also been discussing lately.
The new design should make AMD more competitive in two-socket servers, but the real strategy seems to be to undercut some of these higher-cost Xeon E5 systems with cheaper single-socket Epyc servers. Forrest Norrod, the head of the server business noted that AMD has 0 percent market share in the mainstream two-socket market, so there's nothing to cannibalize with lower-cost systems. Some of these customers currently purchase a two-socket server with only one Xeon E5 chip in order to get certain memory or I/O features. AMD argues that an Epyc single-socket server can provide more cores, the same memory capacity and bandwidth, and more PCI-Express 3.0 lanes than the current Broadwell two-socket systems.
"We have, let's call it very little to lose, and a lot to offer," AMD CEO Lisa Su said. "The datacenter will look very different in a few years."
The company has a relatively modest goal with Naples of getting back to double-digit market share -- though that would still represent a significant increase in units and revenue. But AMD also showed a long-term Epyc roadmap that included Rome, a 7nm chip based on the Zen 2 architecture, and Milan, based on an enhanced 7nm process and the Zen 3 design, while offering few details.
AMD is already shipping the Ryzen 7 and Ryzen 5 processors for desktops, and it said systems based on these chips will be available from Acer, Asus, Dell, HP and Lenovo by the end of the quarter. The company announced a new high-end chip, dubbed Threadripper, which will have 16 cores (32 threads) and greater memory and I/O bandwidth. It will be demonstrated next month at Computex in Taiwan.
Intel is rumored to be preparing a Core i9 processor, which could also be announced at Computex, to compete head-to-head with Threadripper and the fastest Ryzen 7 chips. The issue with these chips is that there are relatively few consumer applications that really take advantage of so many cores and threads.
In the second half of the year, AMD plans to release Ryzen 3 low-end processors as well as Ryzen Mobile APUs (Accelerated Processing Units) that also include new Vega on-die graphics. These will deliver 50 percent better CPU performance, 40 percent better graphics performance and longer battery life than laptops based on AMD's current mobile platform. Ryzen Pro processors for business desktops will also be out later this year, followed by the mobile versions in the first half of 2018.
AMD's current Polaris graphics is targeted at low- and mid-range cards, but the company is expected to release Vega, a new architecture designed to compete head-to-head with Nvidia's fastest GPUs, any day now. The first Vega card, called the Radeon Vega Frontier Edition, will be available in late June and targeted at deep learning rather than high-end gaming -- an indication of just how important this market has rapidly become.
AMD did not provide full specifications, but we know it will have 64 Compute Units and 16GB of HBM2 stacked memory with 480GBps of throughput, and deliver 12.5 teraflops (trillions of floating-point operations per second) at single-precision and 25 teraflops at 16-bit for deep learning applications. That's enough to edge out the Nvidia P100 at 9.3TF single-precision and 18.7TF at half-precision (AMD showed the Frontier Edition besting the P100 on Baidu's DeepBenchBaidu's DeepBench, for example), but it will fall well short of the new V100 announced at last week's GPU Technology Conference, which will deliver 15TF single-precision and up to 120TF at mixed-precision (16- and 32-bit) for deep learning because of its specialized Tensor Cores.
Raja Koduri, the head of the Radeon graphics group, said Vega will be competitive but he was "not declaring victory here over Nvidia or anyone else." AMD also announced a Radeon Pro Vega SSG (Solid-State Graphics) card with both 16GB of high-bandwidth memory and an on-board 2TB NVMe solid-state drive for professional visualization.
Overall AMD emphasized how a more competitive lineup and a predictable roadmap will help it claw back market share from Intel and Nvidia over time. CTO Mark Papermaster talked about how the company was using two "leapfrogging" teams to work on two designs simultaneously and the AMD showed an x86 roadmap, though this didn't reveal much except that there will be a 7nm "Zen 2" and a 7nm+ "Zen 3" with steady improvements in performance per watt by 2020.
On the graphics side, the roadmap includes two generations of Vega (on 14nm and 14nm+) followed by a 7nm Navi family and eventually a "next-generation GPU" on an enhanced 7nm process -- again with no details. AMD did state this it will use both GlobalFoundries and TSMC to manufacture its 7nm chips.