As in just about all of the IT business, things are again changing quickly in the data center. New workloads are taking advantage of advanced technologies, such as artificial intelligence (AI), machine learning and analytics, opening the door for more optimized silicon and pushing back at general-purpose chips.
At the same time, the data center is beginning to move beyond the firewall and into the cloud and edge, where features like power efficiency and throughput are becoming even more important and hyperscale cloud providers like Google and Amazon have the resources to build the chips they need for their workloads.
How does a device manufacturer—and there are a lot more than you might imagine globally—select the perfect processor? It’s not an exact science, but the various chip makers do have their specialties. Intel, the largest processor maker in the world, has a wide selection available; No. 2 AMD is bent on becoming No. 1 and offers some customizations that Intel perhaps cannot; Arm also is aggressive in the market and is the king of low-power, more efficiency-oriented chips; and IBM still cranks out enterprise-capable processors that many customers bet the company on.
Here’s a look at some of the top server chip vendors, with eWEEK’s list being created with help from our own resources alongside analyst firms Gartner, IDC and others. This list is aimed at IT server makers, integrators, component makers; chips are integrated into devices, chipsets or other components in manufacturing processes.
Santa Clara, Calif.
Value proposition for potential buyers: If you’re looking for good, standard general-purpose processors for your IT hardware product, Intel is the place to begin looking.
- Largest chipmaker in the world by volume and revenue, biggest selection and distribution channel available.
- Data Center Group continues to pile up the revenues, garnering $23 billion in 2018, a 21 percent jump over the previous year.
- Readying full production of 10nm chips starting this year, and has unveiled Foveros, a 3D packaging technology that allows for logic blocks to be stacked atop each other and will be the foundation of future chips.
- Intel has expanded beyond its standard Xeon chips. The company now offers field-programmable gate arrays (FPGAs)—increasingly important accelerators—through its acquisition of Altera in 2015 and added additional programmable capabilities by buying eASIC last year. In addition, company officials say that Intel will begin producing discrete GPUs to compete with Nvidia and AMD by 2020.
- Intel has been in business since 1968 and long ago established its market leadership. Chips are considered standards in many use cases.
Advanced Micro Devices
Santa Clara, Calif.
Value proposition for potential buyers: AMD is like Avis—No. 2, but always trying harder to serve customers in order to someday move out of Intel’s shadow. The company is becoming more innovative all the time.
- AMD in 2017 became a serious competitor in the data center with the introduction of its Epyc processors based on Zen, a built-from-the-ground-up x86 microarchitecture. Zen was designed to address the demands for scalability, performance, power efficiency and affordability in modern data centers that are running such workloads as AI, machine learning, virtual and augmented reality and data analytics. It was made with cloud providers in mind.
- Launched the first of its 14nm Zen-based Epyc server chips in 2017, and this year will follow with its 7nm chips, codenamed “Rome.” The new processors will have multiple dies in a single socket, rather than having a single monolithic die, and will include “chiplets”—7nm silicon housed in a 14nm I/O die—to drive compute power as well as fast connectivity. Rome also will have up to 64 cores and 128 threads, double what the current “Naples” Epyc include, and an enhanced version of AMD’s Infinity Fabric.
Arm (owned by Softbank)
Value proposition for potential buyers: Arm is out on the cutting edge of the chip world, innovating some of the finest low-power chips in the world. If you are manufacturing smaller, edge-type products (sensors, videocams, etc.), check ARM first. Its data center processors are also top-fllight.
- Arm doesn’t make chips; instead it designs chips and then licenses those designs to a broad array of manufacturing partners. Its history has been in designing systems-on-a-chip (SoCs) for mobile devices, but company officials for more than a decade have talked about Arm’s low low-power designs being a good fit for data centers that increasingly value energy efficiency as much as performance.
- Marvell—which bought Arm chip maker Cavium last year for $6 billion—is continuing to develop its ThunderX family of chips and startups like Ampere are coming to market. OEMs such as Hewlett Packard Enterprise, Lenovo and Cray are embracing Arm-based SoCs for some systems; AWS has its custom-made Graviton processors for some cloud instances; and Arm SoCs have made their way into supercomputers.
- Arm also now has its own platform for edge and cloud computing environments, dubbed Neoverse. The company unveiled the roadmap in October 2018 and introduced its 16nm “Cosmos” platform, and early this year announced the second generation, the N1 platform, built on the 7nm “Ares” core. It can scale p to 128 cores and 60 percent faster performance than Cosmos. The E1 platform is aimed at high-throughput environments.
Value proposition for potential buyers: If need good enterprise-type processors for servers and other data-center equipment, you may want to inquire here first. IBM’s ecosystem is second only to Intel in providing specialized goods and services.
- IBM for a long time was the king of servers, with its mainframe systems and later its Power-based machines. It also had a line of x86 servers that ran on Intel processors, but sold that business to Lenovo in 2014 for $2.3 billion. With that out of the way, IBM has put its full focus on its Power architecture, and during the past two years has introduced the Power9 chip in a series of powerful servers that offer an option to systems running Intel Xeons.
- Power9 offers more performance per core and almost twice the memory bandwidth performance than Xeon Scalable Processors. At the same time, the company in 2013 helped launch the OpenPower Foundation as a way of expanding the reach of the architecture beyond IBM systems and into servers built by others and running open-source software. As with Arm, partners can license the architecture—IBM opened up such areas as the BIOS and firmware—and last year at the OpenPower Summit there were more than 100 solutions developed with partners like Google, Hitachi, Cavium, Nvidia and Xilinx.
- In addition, Google and Rackspace have a project called Zaiu,s in which they are developing a dual-socket OpenPower-based system designed using Power9. They have contributed the design to the Facebook-led Open Compute Project.
- IBM is the world’s oldest information technology company (since 1916) and has built a reputation of quality for several generations. But server processors are only one of many businesses the company runs.
Marvell Technology Group
Value proposition for potential buyers: Marvell has invested a great deal into its server chip business and is motivated to compete globally with Intel, IBM and Arm.
- Marvell propelled itself into the server chip space last year when it completed a $6 billion acquisition of Cavium, a key player in the Arm server chip market. Cavium had developed the ThunderX line of processors, coming out with ThunderX2 in May 2018.
- Cavium already had run out its first-generation ThunderX SoC, but changed course a bit in 2016 when it bought IP and designs for Broadcom’s Arm-based “Vulcan” chip. Broadcom had just been bought by Avago, which wanted to shed the project, and ThunderX2—which had a soft launch in 2016—was based on the Vulcan designs.
- Cavium rolled out more than three dozen versions of the chip, all optimized for an array of workloads, including servers, storage and networking and ranging from 16 cores to 32 cores. They’re aimed at both data center and cloud workloads, and Marvell now offers a ThunderX2 customer reference platform.
Santa Clara, Calif.
Value proposition for potential buyers: This young, ambitious company has a lot of ground to cover to carve out market share from the big providers, but it has a new-gen approach to servers and cloud services that newer enterprises may find intriguing. New customers and new companies often can mean deep discounts, too.
- Ampere launched in February 2018, founded by a number of ex-Intel executives, including Renee James, who is the startup’s CEO. Company officials said the increasingly distributed nature of the cloud and the requirements of modern workloads like AI round memory and power necessitated a new way of building processors. Ampere is leveraging the work of Applied Micro in the development of its X-Gene Arm-based SoCs. Macom bought Applied Micro in 2017 and sold the X-Gene technology to Ampere.
- Company is targeting hyperscale cloud and edge computing environments. Late last year it announced the Ampere eMAG chips for data centers as well as a roadmap that calls for the next-generation 7nm chip with single- and multi-socket options coming later in 2019.
- The initial chip includes 32 cores running up to 3.3GHz and eight DDR4-2667 memory controllers, and is built in a 16nm FinFET manufacturing process. Lenovo and several original-design manufactures (ODMs) have announced they will use the chip.
Value proposition for potential buyers: Long-established company that has its loyal buyers, mostly from way back. Not known for innovation.
- For a long time, Fujitsu partnered first with Sun Microsystems and then Oracle in developing SPARC chips for servers, going so far as to base its K supercomputer on the architecture. When it launched in 2011, the K system was the fastest supercomputer on the Top500 list. However, company officials announced in 2016 that the follow-up to the K system—which is referred to as Post-K—will be based on custom Arm-based silicon.
- According to Fujitsu, the Post-K supercomputer—powered by the Arm-based A64FX processor and expected to be operation by 2020—will be 100 times faster than K. The chip will leverage Scalable Vector Extension (SVE), a new vector format created with Arm for supercomputers.
eWEEK Score: 4.4/5.0
- Organizations for more than a decade have increasingly relied on accelerators to work with CPUs to improve system performance, keep a lid on power consumption and help run modern workloads like AI and data analytics. It’s market that a report from ResearchandMarkets says will grow from $2.84 billion last year to $21.19 billion by 2023. Nvidia started the accelerator push more than a decade ago with its general-purpose GPUs, and in 2017 rolled out its powerful Tesla V100 GPU. AMD also offers GPU accelerators and Intel will start producing its own discrete GPUs in 2020.
- Demand for FPGAs—which unlike CPUs can be programmed in software and adapted to changing workloads—also has been growing. Xilinx has long been an FPGA vendor, but over the last few years company officials have said that Xilinx is now more of a platform maker for the data center. It not only makes FPGAs but also SoCs under the Zynq brand. Intel, through its acquisition of Altera, also makes FPGAs. ASICs also are becoming key accelerators in servers.
- Hyperscale data center operators like Google, AWS and Facebook all are developing their own silicon to run instances in their cloud operations as well as advanced AI and machine learning workloads. AWS officials in November 2018 announced the company’s Inferentia chips for AI researchers. For now the chips—which will come out later this year—will be focused on the inference part of machine learning, rather than the training. Inferentia was announced around the same time as AWS unveiled Graviton, Arm-based data center processors that will run some of the cloud provider’s instances.
- Other hyperscalers also are developing their own processors for AI workloads. Google in 2016 rolled out its first Tensor chip (TPU) and last year introduced TPU 3.0. Alibaba a year ago unveiled its own AI chips and Intel and Facebook announced in January a plan to develop an AI chip for inference.
The Chinese government has been putting a lot of effort and money into building up the country’s IT industry, with the goal of making China a hub for technology and enable its vendors to build more of their systems with homegrown components, including processors.
- Huawei, the massive tech vendor, in January introduced the Kunpeng 920, an Arm-based SoC that will power the company’s TaiShan servers, which will be aimed at enterprise data centers as well as Huawei’s cloud services offerings.
- In addition, Chinese company Hygon announced last year it was shipping x86 server processors, called Dhyana, that are close in design to AMD’s Epyc chips. It was the result of a joint venture AMD established in the country called Tianjin Haiguang Advanced Technology Investment Co. (THATIC) to license the x86 designs. THATIC was jointly owned by AMD and various Chinese companies, which led to the creation of two other companies, Haiguang Microelectronics and Hygon.
- In addition, the massive Sunway TaihuLight supercomputer, which for two years until last year topped the Top500 list, included a number of Chinese components, including the Sunway Sw26010 processors.