AMD has launched what it called the world's most advanced processor for data centers, featuring up to 64 cores based on the 7-nanometer node and offering double the performance compared to its previous generation of server chips. AMD said its latest EPYC CPUs deliver more than double the performance per dollar of Intel's Xeon CPUs, giving it an opportunity to gain ground in the highly profitable server chip business.
AMD said its second-generation EPYC CPUs—code named Rome—doubles the performance and cuts the power consumption of its first generation of server chips—code named Naples. The new processors offer more than 50% better performance and up to 40% lower costs than Intel's latest Cascade Lake CPUs, AMD said. That reduces costs for cloud and enterprise customers deploying the chips in data centers by 25% to 50%.
Lisa Su, AMD's chief executive officer, said in a statement that its second-generation server chips "set a new standard" for the data center segment. She added that its latest EPYC CPUs are the highest performance x86 processors in the world, taking the throne from rival Intel. Shares of AMD have surged around 17.5% since the announcement of the new EPYC CPUs, which are based on the same Zen 2 architecture as its Ryzen CPUs for PCs.
The Santa Clara, California-based company has been on a crusade against Intel’s stronghold in data centers, where the world’s largest chip manufacturer holds more than 95% market share, which has cooled significantly after last year's spending surge. Since Su became chief executive in 2014, AMD has been out to challenge Intel's Xeon CPUs—the current gold standard in private and public cloud data centers—not only on price but also on performance.
AMD is also trying to take advantage of the delays in Intel's development of server chips using its 10-nanometer node. AMD's latest line of server chips is based on the 7-nanometer node, the most advanced and most costly chip production technology from TSMC, raising the stakes in AMD's battle against Intel. Using the 7-nanometer node gives AMD's EPYC CPUs higher core counts than Intel's Cascade Lake Xeon CPUs, which are based on 14-nanometers.
Intel plans to start shipping the last generation of server chips based on 14-nanometers, called Cooper Lake, in the first half of 2020. AMD's Silicon Valley rival has pushed out production of server chips based on the 10-nanometer node—which is considered roughly as advanced as the 7-nanometer node from TSMC—to the second half of 2020, giving AMD an opportunity to win over customers from Intel, which is struggling to get through its manufacturing glitches.
AMD said sales of its Rome CPUs are expected to expand "significantly faster" than its Naples CPUs, which started shipping over a year ago. The company's server chips have started to gain speed among major cloud vendors, more than doubling its market share over the last year to 3.4% at the end of the second quarter of 2019, according to Mercury Research. “Adoption of our new leadership server processors is accelerating," Su said in a statement.
AMD is aiming to reach double-digit market share in server chips before Intel starts selling its 10-nanometer Ice Lake CPUs in the second half of 2020. Google said that it had started using AMD's Rome CPUs in its data centers and plans to offer them to developers over the cloud at cheaper costs than Intel's Xeon CPUs. AMD has courted other major cloud vendors, including Amazon, Alibaba, Baidu and Microsoft, all of which are big buyers of Intel's server chips.
"AMD needs to aggressively show that it has a compelling roadmap and step on the gas for the next year to steal as much share as it can," Patrick Moorhead, principal analyst at Moor Insights & Strategy, said in a research note. Every private enterprise, server manufacturer and cloud vendor "wants more competition in the space to accelerate innovation and lower costs. With that said, none of these customers would adopt AMD if it didn't have some advantages."
Moorhead said AMD could win up 10% market share with its new EPYC processors.
The second-generation server chips are based on chiplets connected through AMD’s Infinity Fabric interconnect, which boosts the bandwidth between them to act as though they share the same die. The new EPYCs integrate up to eight CPUs based on the 7-nanometer node from TSMC surrounding the I/O core produced on 14-nanometers from Globalfoundries. The I/O takes care of communications to memory and accelerators such as AMD’s Radeon GPUs.
Building its server processors out of smaller die based on 7-nanometer production enables AMD to accommodate more cores at the same power consumption. AMD said the Rome CPUs feature up to 64 cores, doubling the core count of its 32-core Naples CPUs and outclassing Intel’s 56-core Cascade Lake Xeon processors. Intel plans to put up to 56 cores in its upcoming Cooper Lake CPUs, which are expected to enter mass production in the first half of 2020.
AMD has added other advanced technology to steal market share from Intel's server processors, which can cost more than $10,000 each on the high end. While Cascade Lake CPUs support the PCIe Gen 3 standard for connecting to additional storage and accelerators cards for artificial intelligence, AMD's new EPYC CPU supports 128 lanes of PCIe Gen 4 connectivity, doubling the bandwidth for cloud, enterprise and high-performance computing customers.
AMD's latest line of chips range from 8 cores to 64 cores, with the base frequency of the chips running from 2 GHz to 3.2 GHz. The flagship EPYC 7742 features 64 cores, 128-threads and 256 MB of shared cache and costs $6,950, while the base EPYC CPU contains 8 cores, 16 threads and 32 MB of shared cache and costs $450, according to AMD. The memory controller inside every chip can access up to 4 TB of memory at a maximum speed of 3.2 GHz, AMD said.
AMD said that the top-of-the-line EPYC processor boasts 80% better performance for 40% of the price of a high-end Cascade Lake CPU—the Xeon 8280L—which is based on Intel's 14-nanometer node and contains 28 cores. The chip's base frequency is 2.25 GHz with power consumption of 225 W. The 64-core Rome CPU also pumps out 23% more instructions per cycle (IPC) than its 32-core Naples CPU. AMD has also enlarged the L3 cache from 64 MB to 256 MB.
Intel has been trying to repel rivals in the data center business by selling a broader portfolio of products. The Silicon Valley company has started offering its Optane memory to customers buying its Cascade Lake CPUs, which have been enhanced to handle artificial intelligence. Intel has also doubled down on advanced packaging technology that can be used to create more affordable custom chips by mixing and matching CPUs, GPUs, FPGAs, ASICs and other chiplets.
But challenging Intel's dominance in data centers is a David versus Goliath battle. Intel's sales of server chips came to $9.9 billion in the first half of 2019, more than AMD's total revenue of $6.5 billion in 2018. But Intel has not only surrendered its manufacturing lead to TSMC over the last year but has also started to lose its momentum to AMD. Shares of AMD have added 80% since the start of the year, to $34.20. Intel's shares have shed more than 2% in 2019.
AMD has been slowed down by speed bumps. Sales in the second quarter slipped by 13% over the last year to $1.53 billion as revenue from semi-custom chips that power gaming consoles such as Sony's Playstation cratered. The company's profits of $35 million in the second quarter dropped 70% from the same quarter last year. Last month, AMD was also forced to cut its forecast for 2019 sales growth from the high-single digits to a mid-single digit percent.
But the company is still in position to plunder market share from Intel in the second half of 2019. Profit margins came out to 41% in the second quarter, up from 37% from a year ago. AMD sees gross margins growing to 43% in the current quarter spurred by buoyant sales of its 7-nanometer Ryzen CPUs, 7-nanometer Radeon GPUs and 7-nanometer EPYC CPUs. In addition, AMD expects its 2019 revenue growth to reach 20% excluding sales of semi-custom chips.