Microsoft, Google Gravitate to AMD’s Rome

Article By : Rick Merritt

Google and Microsoft said they will use Rome, AMD's latest x86 processors battling for Intel's lucrative server market.

AMD is likely to claw more sockets back from Intel despite an uncertain server market with its second-generation Epyc launched today. The combination of the chip’s upgraded Zen x86 cores made in a 7nm TSMC process along with more PCIe and DRAM channels deliver more performance and I/O at less cost than Intel’s latest Xeon.

Analysts said AMD’s Epyc 7002 series, aka Rome, likely will meet its goal of winning the company at least 10% of classic server sockets by next June, up from its low- to mid-single digit share today. To date, most of the gains AMD has made with its new Zen-based processors have been in desktops and notebooks.

The server remains the hardest and most lucrative nut to crack, and AMD showed progress at an event here. Google Cloud, Microsoft Azure and Twitter are early users. Such hyperscalers are likely to be its main customers in the short term, given AMD’s relatively small ecosystem compared to Intel in business computing software, one analyst said.

AMD has yet to make a major break into Intel's dominance in servers, according to figures from Mercury Research.
AMD has yet to make a major break into Intel’s dominance in servers, according to figures from Mercury Research.

The 7002 Series integrates up to 64 dual-threaded Zen-2 cores running at a base frequency of up to 3.1 GHz with up to 256 Mbytes L3 cache. Each processor’s package contains a 14nm I/O hub supporting 8 DDR4-3200 DRAM channels and 128 PCIe Gen 4 links.

Intel’s latest Cascade Lake processors sport fewer cores, DRAM channels and PCIe Gen 3 links. As a result, AMD claimed the 7200 delivers 80 new performance records, an average of up to twice the performance, and more than twice the performance/dollar of Cascade Lakes.

“Rome is the highest performance x86 processor in the world…Epyc will be the new standard for the modern data center,” said Lisa Su, AMD’s chief executive.

Thanks to its system-in-package approach, AMD packs nearly 1,000mm2 of silicon and 32 billion transistors in a device. The 7002 uses an upgraded Infinity Fabric running up to 18 GTransfers/s to link the die. The new design flattens the memory hierarchy and lowers overall latency from AMD’s first generation Epyc, called Naples.

Google is already using Rome in its data centers and will offer it as part of its public cloud service later this year. Bart Sano, a vice president of platform engineering at Google praised the CPU for its “high core and thread count, its shared memory, and PCIe Gen 4 that helps us scale and integrate our own accelerators efficiently,” he said, presumably referring to its TPU.

Microsoft will use Rome in three of its Azure cloud services. They include a next-gen instance for high performance computing (HPC) leveraging the PCIe links to support native Infiniband instance, a virtual desktop service also using AMD’s Radeon GPUs, and two-general purpose systems currently available for review.

Rome delivers twice the performance on HPC workloads as Naples, said Girish Bablani, a vice president for the Azure compute service.

An engineering manager from Twitter said it will use Rome in its production networks before the end of the year. The chip packs 40% more cores in a rack than Intel processors, lowering total costs by 25%, she said.


The Epyc 7002 Series spans 120W to 225W processors (Source: AMD)

On OEMs, road maps and market shares

Among OEMs, Hewlett-Packard Enterprise said it has three Rome systems available today and will have 12 within a year. Lenovo announced two Rome systems, in part targeting users who need lots of solid-state drives. Dell also plans to use Rome in servers shipping this fall but did not announce details.

For its part, Cray is “approaching $1 billion in new bookings” for its Shasta HPC systems using Epyc, said chief executive Pete Ungaro. The company recently landed deals for three government systems based on Rome and it plans to build an exascale system probably using the 2020 version of Epyc.

AMD said it has already completed the design of a next-generation Zen-3 x86 core to be made in TSMC’s N7+ process for a server chip called Milan, scheduled to ship next year. It is now designing a Zen-4 core for a follow-on server processor called Genoa.

Over the past year, the Naples chip brought AMD from about 1.4% to about 3.4% of the broad server market, according to Dean McCarron of Mercury Research. Other analysts said AMD should hit its target of commanding 10% of the server market by June 2020.


A combination of 7nm TSMC cores and a 14nm Globalfoundries I/O chip keep Epyc’s costs low. (Source: AMD)

Naples’ had relatively slow single-core performance and was more power hungry, problems Rome fixed, said Linley Gwennap of the Linley Group. Intel likely will have an edge in inference jobs given its DL Boost instructions and will shine with in-memory database tasks with its Optane DIMMs, said Patrick Moorhead of Moor Insights & Strategy.

Moorhead sees AMD gaining ground with the hyperscalers but less so in companies’ private data centers. “AMD is investing a lot more in the enterprise over the last two years but nowhere to the degree Intel has been investing the last decade. [So,] AMD will need to lean on OEMS like HPE, Dell and Lenovo, none of whom have a recent track record of creating demand for AMD,” he said.

AMD targets 10% of a market of about nine million classic servers a year while Intel pursues a broader market of 12-14 million units/year that include storage and networking systems, said Shane Rau, an analyst for International Data Corp. The server market will be flat to 5% up this year depending on whether hyperscalers resume buying at something like historical rates this fall.

In a research note, Jeffrey Kvaal, a financial analyst at Normura Instanet, said capital spending in the U.S. from Web giants hit $24.3 billion in the second quarter. He called it “a return to double-digit year-over-year growth after a 5% decline in 1Q” though still 9% below forecasts.

After a rush of spending on servers and other gear over the last two years, hyperscalers pulled back in the first quarter. McCarron of Mercury called the resulting decline for server CPUs “the worst in a decade due to low cloud demand and a very steep drop in enterprise/government business.”

Leave a comment