Qualcomm has announced the commercial shipment of Centriq 2400, the server processor range on which it is basing its attack on Intel’s heartland in the cloud infrastructure market. It was boosted by endorsements from three of the webscale giants which will be key targets for new server architectures – Microsoft Azure, Google and Alibaba (though no confirmed deals, and an absence of Amazon or Facebook).
The system-on-chip, announced last year, is the first server processor to be manufactured in the 10 nanometer process and claims to be the highest performance ARM-based server chip, as it takes on not just Intel, but more established vendors of data center and cloud processors based on the ARM architecture, notably Cavium.
“Today’s announcement is an important achievement and the culmination of more than four years of intense design, development and ecosystem enablement effort,” said Anand Chandrasekher, the former Intel executive who now heads up Qualcomm’s new Datacenter Technologies unit. “We have designed the most advanced ARM-based server processor in the world that delivers high performance coupled with the highest energy efficiency, enabling our customers to realize significant cost savings.”
The chip is manufactured by Samsung using its 10nm FinFET process and runs on 48 single-thread, 64-bit cores, running at up to 2.6 GHz frequency. The cores, called Falkor, are ARM-based but customized by Qualcomm under its architectural licence. The cores are connected with a bidirectional segmented ring bus with 250GBps of aggregate bandwidth, and there is 512KB of shared L2 cache for every two cores, and 60MB of unified L3 cache distributed across the die.
It emphasizes the aspects of the ARM architecture which are attracting the interest of server makers – low power consumption and low cost. Centriq 2400 claims to consume less than 120W of power and costs $1,995, which Qualcomm says represents four times better performance per dollar, and up to 45% better performance per Watt, than Intel’s top end Skylake processor, the Xeon Platinum 8180.
“We brought this power efficiency from the mobile space into the data center space and that’s something that the world needs,” said Paul Jacobs, executive chairman of Qualcomm, at the launch.
Centriq 2400 also supports ARM’s TrustZone security operating environment, and hypervisors for virtualization.
Chip analysts were generally positive. “It should be competitive. There is a lot of interest in ARM server processors because no two workloads are the same,” Jim McGregor, principal analyst at semiconductor watchers Tirias Research, told IDG News, indicating that Centriq would be particularly suited to cloud workloads, web servers and enterprise data centers doing different instantiations. But it will probably be less suited than Intel to workloads with huge data sets.
Of course, a disadvantage is that the huge base of applications for Intel’s x86 architecture must be rewritten for an ARM platform, but this issue is becoming less serious as many enterprises and cloud providers shift towards very flexible software based on virtual machines and microservices. That is making them more willing to use different architectures for different workloads.
Jacobs said: “In the traditional IT environment apps and services are delivered through scaling up with bigger servers but in the cloud environment they are delivered through scaling out – more and more servers … That has really changed the design point for the server – it creates this tremendous opportunity for Qualcomm.”
These are the messages potential customers will want to hear, and Qualcomm – like Cavium and other ARM supporters in the infrastructure market – is likely to focus initially on low hanging fruit, such as Cloud-RAN systems for its traditional user base of mobile operators; but also to lust after the prize customers, the webscale cloud operators such as Google, Facebook, Amazon and Microsoft, with their huge deployments of servers. All these companies are reported to be experimenting with ARM-based architectures, though they may choose to implement their own inhouse designs rather than buying off-the-shelf – a threat which is pushing Intel, Qualcomm and others into higher levels of customization.
Qualcomm ticked many of their boxes in its claims for its new offering, saying it was “optimized for common cloud workloads … designed to deliver throughput performance for highly threaded cloud native applications that are developed as microservices and deployed for scale-out.”
The vendor did manage to field a decent array of cloud-oriented companies for the launch event, many of them demonstrating applications on Centriq, though commercial deals were not guaranteed. They included Chinese ecommerce giant Alibaba, Google and Microsoft Azure. Amid the various partner quotes with which Qualcomm adorned its press release, three will have been particularly welcome, coming from those cloud giants.
“Alibaba has been collaborating with Qualcomm Datacenter Technologies for several years on our need for a high performance and power-efficient server chip for our cloud solutions,” said Weifeng Zhang, senior director of Alibaba Infrastructure Services. “We see great growth opportunities for the Qualcomm Centriq 2400 processor family in the burgeoning ARM-based server ecosystem in China.”
“Google is excited to see Qualcomm Datacenter Technologies launch the Qualcomm Centriq 2400 processor,” said Bart Sano, VP of platforms at Google. “We welcome choice in the processor design space for data centers. Choice leads to innovation which ultimately benefits our users. The 64-bit ARMv8-A architecture and ecosystem is now a viable alternative for scale-out data center designs.”
“Microsoft and Qualcomm Datacenter Technologies have made tremendous progress since our announcement at the Open Compute Project US Summit earlier this year,” said Mike Neil, corporate VP of Azure infrastructure and management. “We believe that it has the ability to deliver extremely compelling performance per watt and total cost of ownership.”
But Qualcomm will not have a clear run at these critical customers. Earlier this year, Microsoft announced that Cavium was to be a participant in its Project Olympus open architecture server initiative. Other ARM server platforms include The X-Gene 3 server platform developed by AppliedMicro (now part of MACOM); while AMD offers an x86-based alternative, the Epyc, which has been positively received.
While Qualcomm has been involved in trials of its architecture with Google, in March, the search giant took the stage at the OpenPower Summit, which was taken to signal that it was favouring that design over an ARM one. At the time, Google said it had built software which allows it to switch easily between Intel and other platforms – and used the IBM-led OpenPower design as its example.
Among Qualcomm’s significant advantages over some other ARM-based server challengers is its understanding of low power, high volume platforms, thanks to its mobile heritage. And the scale of its mobile business also gives it a high degree of influence over its manufacturing partners, and the ability to gain early access to the most cutting edge process nodes, including 10nm.
Of course, no company which buys as many processors as Google (or Facebook or Microsoft) can afford to limit its supplier choices and the cloud giants take an interest in almost any second source to Intel, even investing in their own inhouse design projects. If the ARM server platform starts to gain real scale – and Qualcomm is the most likely driver of that – it will certainly be of interest to the webscale majors.
Google, along with the other operators of hyperscale cloud data centers like Amazon, Facebook and Microsoft, has the power to break Intel’s still-intact dominance of the server space. While x86 will go unchallenged for a long time in very high performance computers and most enterprise servers, in the high growth area of low power cloud servers, as well as emerging categories of infrastructure like Cloud-RAN platforms, there is more space for disruption from an architecture promising new levels of power and cost efficiency – and an end to the x86 hegemony.
Google is one of Intel’s biggest customers, so the loss of even some of its business would be a severe blow to a server business which has been performing strongly. Google’s parent firm, Alphabet has almost trebled its infrastructure spending since 2011, and other companies driving server growth, like Amazon, are also looking for ways to reduce the vast cost of servicing global web, cloud and data usage. Google buys an estimated 1.2m server processors a year, or 5% of the total market, according to IDC, with only Hewlett-Packard Enterprise and Dell purchasing more.
Qualcomm is certainly determined to take some of Intel’s playground for itself. At the start of this year, it announced a joint venture with the hi-tech Guizhou Province in China, focused on developing and selling its new server processors in that market. The US firm estimates that server silicon will be a $15bn addressable market for it by 2020, with $6bn of that potential located in China. Baidu, the ‘Chinese Google’ is already testing ARMv8 processors and will be a hot target for the Qualcomm venture, which recently started work on China-specific implementations of Centriq.
However, there have been many predictions of a strong ARM server market, and so far this has not materialized. Early attempts suffered from being confined to 32-bit, though that was remedied with ARMv8. The 10nm core is an important breakthrough, which should help an ARM design, for the first time, compete with Intel Xeon on performance, not relying on power efficiency alone to score points. A 10nm Xeon is not expected until next year.
It looks as though 2018 will be the decisive year, because demand from telcos for virtualized infrastructure will achieve significant scale in that year, after a slow 2017; and because there will be several new and powerful entrants. As well as 10nm Xeon, Cavium is expected to ship its 14nm Thunder X2 with 54 cores running at 2.6 GHz, making it a real workhorse for applications like Cloud-RAN. And the first non-IBM OpenPower server processors are expected to emerge from China then, opening up the choice of suppliers and the ecosystem around that architecture.
Other supporters at the Centriq event:
Server manufacturer HPE was present at the Centriq launch, as were chip partners like Mellanox and Qualcomm’s highly strategic FPGA ally, Xilinx (FPGAs, with their programmability, are becoming increasingly important in the network and cloud infrastructure world to add a dose of flexibility that microprocessors cannot promise – hence Intel’s acquisition of Xilinx’s arch-rival Altera).
Other attendees included web services and development tools providers such as LinkedIn, Canonical, Red Hat, SuSE and Uber. Some of the applications which were demonstrated included:
- Web front end with HipHop Virtual Machine
- NoSQL databases including MongoDB, Varnish, Scylladb
- Cloud orchestration and automation including Kubernetes, Docker, metal-as-a-service
- Data analytics including Apache Spark
- Deep learning inference
- Network function virtualization
- Video and image processing acceleration
- Multi-core electronic design automation
- High throughput compute bioinformatics
- Neural class networks
- OpenStack Platform
- Scaleout Server SAN with NVMe
- Server-based network offload