AMD Introduces Epyc Server Chip, 16-Core Desktop Chip

amd-introduces-epyc-server-chip-16core-desktop-chip photo 1

At this week's financial analyst meeting, AMD unveiled a 16-core, 32-thread desktop processor called Ryzen Threadripper—its new Epyc brand for server chips—and introduced its first graphics board aimed at the machine learning market.

But I was also glad to see the company unveil a roadmap of successive generations in its CPU, GPU, and server lines, with migration to 7nm and 7nm+ process nodes through 2020, which is crucial for the firm to regain credibility with business buyers. At this point, competitors Intel and Nvidia dominate their markets, particularly on the server side, and business buyers need to be convinced that AMD will be a long-term player in order for it to land on the consideration list.

"Immersive and instinctive computing will transform all of our daily lives," AMD CEO Lisa Su said at the conference when defining the company's vision for the future. Immersive computing with high-end graphics is all around us, but instinctive computing—which involves using huge amounts of data and machine learning algorithms—is just beginning to evolve. All of this requires "high-performance computing," she said, a term she used to describe all sorts of high-end computing and graphics, not just the HPC or supercomputing market.

Su talked at length about investing for "multi-generational leadership" in x86 CPUs, graphics for both PCs and integrated, and software—a big change for a company whose primary products have targeted the mainstream or low-end markets. AMD will now focus on premium products, she said, and noted that while the mainstream accounts for most of its units, the premium part of the market accounts for most of its revenue and profits.

amd-introduces-epyc-server-chip-16core-desktop-chip photo 2

Her biggest announcement was probably Epyc, the branding for the new line of server chips, which had been codenamed Naples. Su said that the Zen architecture, which debuted in the Ryzen desktop chips was "created with the new data center in mind." And while she is enthusiastic for what Zen can bring in the desktop and laptop markets, she is even more excited about what Zen can do in the data center market. She talked about how AMD has also targeted the data center market through Radeon Instinct, a version its next-generation graphics architecture known as Vega—which will provide 25 teraflops of performance—and how these would work together in a vision of heterogeneous computing.

"Today's data center really requires heterogenous computing to be successful," she said, describing AMD as the only provider of both high-performance computing and graphics.

amd-introduces-epyc-server-chip-16core-desktop-chip photo 3

CTO Mark Papermaster gave more detail about the glue that will hold the new chips together, as well as the firm's process for designing new chips in a way that will "provide sustained innovation going forward." Papermaster described the major features of the Zen and Vega architectures for CPUs and GPUs, most of which had been described previously, and said the firm has to design "not only for performance, but also for efficiency."

amd-introduces-epyc-server-chip-16core-desktop-chip photo 4

Tying the chips together is the firm's new Infinity Fabric, which connects CPUs, GPUs, memory controllers, and other features within a chip and between chip sockets. Papermaster called Infinity Fabric a "hidden gem," and explained how it includes a control fabric that manages sensors on the chip; can regulate performance and security; and can also work as a data fabric, moving data between the various parts of the system.

He said this allows for "near perfect scalability" through 64 CPU cores. AMD is also supporting some new industry standard interconnects between systems, known as Gen-Z & CCIX, as it pushes for open standards (since it doesn't have the heft of Intel in the market).

Papermaster said that one big challenge for the firm over the next few years will be "defying the slowing of Moore's Law." He said that through integration, software, and system design it would be able to stay at the pace of generational performance improvement, even without frequency improvements.

amd-introduces-epyc-server-chip-16core-desktop-chip photo 5

amd-introduces-epyc-server-chip-16core-desktop-chip photo 6

To that end, he showed roadmaps for both the graphics and CPU lines through 2020, saying that the team was not only rolling out the current generation of products, but working on the next two. The plans show moving to 7nm and 7nm+ production modes, with continuing improvements in both raw performance and performance per watt. While these weren't detailed, they were great to see.

A 16-core, 32-Thread CPU and a GPU Aimed at Machine Learning

amd-introduces-epyc-server-chip-16core-desktop-chip photo 7

Jim Anderson, General Manager of the Computing and Graphics group, said that in addition to the Ryzen 7 and Ryzen 5 products already announced, a lower-end Ryzen 3 will ship in the third quarter. More importantly, all five of the top PC OEMs (Acer, Asus, Dell, HP, and Lenovo) will have Ryzen desktops available for consumers in the market by the end of the quarter. Commercial desktops should follow in the second half of this year, he said.

For those interested in the "absolute highest performance," Anderson announced a new version called Ryzen Threadripper with 16 cores and 32 threads coming "this summer."

Anderson said that a mobile version of the Ryzen processor, with on-die Vega graphics, will be available for consumer systems in the second half of 2017. A commercial version should follow in the first half of 2018. He said the mobile chip will offer 50 percent more CPU performance and 40 percent more GPU performance while using 50 percent less power compared to the company's current seventh-generation APU.

amd-introduces-epyc-server-chip-16core-desktop-chip photo 8

Meanwhile, in the graphics world, Raja Koduri, Chief Architect for the Radeon Technologies Group, discussed the company's plans for a new series of graphics boards based on the new Vega architecture. Koduri noted that the company's current Polaris architecture GPUs mostly address the mainstream and mid-market segments of the graphics market (with graphics boards under $300), and acknowledged that the company didn't play at the top end.

This will change with the new Vega architecture. Among the features Koduri described were a new high-bandwidth cache controller (which can double or quadruple the available memory), a new programmable geometry pipeline, Rapid Packed Math (for 16-bit floating point), and an advanced pixel engine. He showed demos featuring much smoother motion in high-end gaming and said Vega will support 4K 60 Hz gaming.

For the professional market, Koduri talked about getting more certifications and supporting SSG, which will allow for a 16GB high-bandwidth cache and up to 2TB of on-board SSG NVME flash memory. He showed demonstrations of this working in real-time ray-tracing and on producing 8K video clips in Adobe Premiere.

amd-introduces-epyc-server-chip-16core-desktop-chip photo 9

Koduri then turned his attention to machine learning, where competitor Nvidia has made huge inroads in deep learning with its GPU-based products and its CUDA architecture. He acknowledged that "AMD is not even in the conversation today," but emphasized the company's heterogenous computing approach and what it calls the Radeon Open Compute Platform (ROCm), which supports running machine learning applications on all of the leading frameworks, including TensorFlow and Caffe.

Koduri demonstrated the new product, which will be called Radeon Vega Frontier Edition, and showed that it is very competitive on Baidu's DeepBench machine learning training benchmark. He said this will offer 13 teraflops of performance at 32-bit, 25 teraflops at 16-bit, as well as up to 16GB of high-bandwidth memory (HBM2), which is four times that of AMD's current Fury X board. This is due out in late June. One would assume other Vega-based boards will follow shortly.

Koduri said that he believes Threadripper and Naples will be disruptive, in part because the CPU bottlenecks will go away, leaving more room for GPU performance.

AMD Promises 'A New Day for the Datacenter'

Forrest Norrod, SVP and General Manager for the Enterprise, Embedded and Semi-Custom Business Group, gave details on the Epyc server ship. He said "datacenter leadership" is the company's No. 1 priority, despite having what he described as "a rounding to zero percentage market share."

Norrod said the company's previous achievements in server technology—which include having created the first 64-bit x86 cores, high-speed coherent interconnects, and integrated memory controllers—are what make it "plausible… predictable… inevitable" for it to again compete in this market.

amd-introduces-epyc-server-chip-16core-desktop-chip photo 10

Norrod said that Epyc will include 32 Zen cores, 8 memory channels, 128 lanes of high bandwidth I/O, and a dedicated security engine. But while that sounds like a huge chip, Norrod explained that the Infinity Fabric Papermaster described earlier enables the chip to actually be built from four 8-core die, which makes it much easier and less expensive to produce.

Norrod said that in a two-socket configuration, the chip could offer 64 cores, 4TB memory, and 128 PCI Express lanes, thus giving it 45 percent more cores, 122 percent more memory bandwidth, and 60 percent more I/O than the Xeon E5-2699A v4 (Broadwell). And he presented a couple of demos that showed it outperforming the Xeon in tasks such as compiling Linux.

I was more impressed with a demo of how a single-socket version of the chip could outperform a middle of the market dual-socket Intel configuration, which he said accounts for most of the market. (As always, I take vendor benchmarks with a grain of salt, and suggest you do as well.)

Epyc will offer the "best value for end users," he said. AMD is looking for leadership in specific segments of the market; Norrod noted it's likely to get traction at first from the bigger datacenter customers who write their own software. One big difference between AMD and Intel's approach, he said, is that every Epyc will be "unrestrained, with all of the I/0, memory channel, high-seed memory, security stack, and integrated chipset features supported on all models. (Intel offers certain features only on higher-end models).

amd-introduces-epyc-server-chip-16core-desktop-chip photo 11

Norrod said Epyc offers a simpler architecture compared to Intel, and made a big deal about combining Epyc and the Radeon Instinct platform for machine learning. Norrod said Epyc is scheduled to be launched in late June, with more than 30 server models expected to be shipped this year.

amd-introduces-epyc-server-chip-16core-desktop-chip photo 12

Norrod also stressed that this is the first in a series of chips, and presented a roadmap with versions called "Rome" and "Milan" (to follow "Naples") between now and 2020, using 7nm and 7nm+ processes. He emphasized that it isn't all about core performance, but rather continuous innovation.

Following a financial presentation, CEO Lisa Su returned to close the event, and told the financial analysts that while the numbers are important, "this company is all about the products." It's certainly good for the industry to have more competition in CPUs, graphics, and especially in the data center market. After all, the periods when we've seen the most competition have also been the periods when we've seen the most innovation.

Recommended stories

More stories