• United States

AMD introduces Zen 3 architecture and pursues Xilinx acquisition

News Analysis
Oct 13, 20204 mins
Computers and PeripheralsData Center

AMD's new Zen 3 processor architecture is socket-compatible with older designs but with faster performance.

AMD had a busy week last week. It introduced the third generation of its Zen microarchitecture, which has been propelling the company’s comeback since 2017, and is the subject of reports it is looking to buy field-programmagle gate array (FPGA) maker Xilinx.

Five years ago, AMD was a non-entity in the CPU market and only kept afloat by its GPU business. Intel had written the company off and considered Qualcomm its biggest competitor. Then the company came out with Zen, a whole new design. “We started with Zen from scratch, starting from a clean sheet of paper,” said CEO Lisa Su in a video announcement.

The result is a nice comeback for a company that had been written off five years ago. It has 5.8% of the server market share as of Q2, 19.2% of desktop and 19.9% of mobile, according to Mercury Research, which specializes in semiconductor market share. The server share may seem low, but two years ago it was at zero and server turnover is slower than desktop.

The new Zen 3 architecture is based on a seven-nanometer design and comes with a 19% uplift in instructions per clock (IPC) over the previous-generation Zen 2. It also has increased maximum boost frequencies, and a redesigned core layout and cache topology. As with previous generations, Zen 3 will be used in AMD’s Ryzen desktop and laptop CPUs and Epyc server processors, just with different core counts and caches.

This announcement was geared around Ryzen and gaming in particular, but there will eventually be new Epyc server processors based on Zen 3. Zen 2 introduced some major changes, notably the chiplet design. Instead of one monolithic CPU, it had several smaller chips with eight or 16 cores per chiplet, all interconnected by a high-speed interconnect.

AMD CTO Mark Papermaster said Zen 3 increased integer and floating point execution units wider and more flexible, and allows for more execution capability at a lower latency. AMD added more branch prediction to reduce delays of computation. That will apply to Epyc even though AMD is currently talking games.

At the same time AMD worked on power efficiency, something that is definitely an issue. Papermaster said the new generation of chips are up to 24% more power efficient than the last generation of Zen, something that will no doubt be appealing to server customers.

Zen 3 Epyc processors, being developed under the codename “Milan,” are planned for launch later this year, AMD says.

How buying Xilinx could help AMD

Last week the Wall Street Journal reported AMD was in advanced talks to buy FPGA chip maker Xilinx in a deal that could be valued at more than $30 billion. That’s the biggest purchase this side of Nvidia’s $40 billion buyout of Arm Holdings.

That’s a hefty sum to pay for a company that might break the $8 billion mark in annual sales this year, but AMD’s comeback has seen a massive run-up in its stock. AMD stock once traded at $0.50. It closed on Friday at $83.10.

Xilinx is in San Jose, California, just down the road from AMD in Santa Clara. It’s the last major independent FPGA maker after Intel gobbled up its rival Altera in 2015 for $16.7 billion. That hasn’t proved a big boom for Intel, notes Kevin Krewell, principal analyst with Tirias Research.

“Intel was trying to create a one-stop shop for AI accelerators and Altera was part of that solution, but it hasn’t been a big win for Intel,” he said. “Altera is still holding the same marketshare.”

But he notes Intel has multiple processor offerings for AI, including Xeon, the Xe GPU, Nervana, and other products. Its strategy is built around an API that uses all of these chips for AI processing. AMD’s AI strategy has been wholly reliant on its Radeon GPUs.

“This would give AMD more flexibility, and Xilinx has gone much further down the road on AI. It would definitely improve AMD’s AI positioning, whereas with Intel Altera was one slice of a bigger pie,” said Krewell.

Andy Patrizio is a freelance journalist based in southern California who has covered the computer industry for 20 years and has built every x86 PC he’s ever owned, laptops not included.

The opinions expressed in this blog are those of the author and do not necessarily represent those of ITworld, Network World, its parent, subsidiary or affiliated companies.

More from this author