eWEEK content and product recommendations are editorially independent. We may make money when you click on links to our partners. Learn More.
IBM officials are rolling out a new Power server that they say is more powerful and efficient than competing Intel-based systems when running such compute-intensive workloads, such as artificial intelligence, deep learning and data analytics.
The new Power System S822LC is the latest server in the company’s Linux server portfolio and features innovations developed through the IBM-led OpenPower Foundation. The system runs on revamped Power8 chips and will give organizations an alternative to servers powered by Intel’s x86-based Xeon processors, according to IBM officials.
The key innovation with the S822LC is new interconnect technology from Nvidia embedded in the Power8 chip that officials said will enable data to flow between the CPU and graphics chips five times faster than in x86-servers. The system embeds Nvidia’s NVLink high-speed interconnect technology into the silicon to connect the Power8 chips with Nvidia’s new Tesla P100 Pascal GPU accelerators.
The result is that the S882LC can run data-intensive workloads significantly faster than x86 systems, and fewer systems are needed in the data center. According to Dylan Boday, an IBM Power engineer, Chinese internet service provider Tencent has been testing the new Power8 systems and finding that workloads run three times faster and that it can reduce the number of servers in its data centers by two-thirds.
NVLink is a critical part of the S822LC system, Boday told eWEEK. The link communication between the CPU and GPU has always been a choke point in servers that use accelerators, he said. With NVLink, “that no longer is a bottleneck.”
Also on Sept. 8, IBM officials introduced two other Power8 Linux servers—the S821LC and S822LC for big data—that also take advantage of Nvidia GPU accelerators, though they don’t use the NVLink technology. Instead, the Tesla K80 GPUs are attached via PCIe. In addition, users can employ IBM’s PowerAccel Coherent Acceleration Processor Interface (CAPI) as a high-speed interconnect with field-programmable gate array (FPGA) accelerators.
IBM sold its x86 server business in 2014 to Lenovo for $2.1 billion in an effort to put more resources into its Power architecture. Intel chips run in more than 90 percent of all servers worldwide, and analysts have said that businesses are looking for other silicon options to help drive innovation, fuel competition and keep chip prices down. IBM and ARM are pushing to become that alternative.
Over the past couple of years, IBM officials have pushed to grow the capabilities of open technologies on the Power architecture. That includes expanding support for Linux and leveraging partnerships through the OpenPower Foundation, which was launched to enable other vendors to develop products and technologies off the architecture and to accelerate innovation.
“It’s is a collaborative environment—a lot of groups are innovating around Power,” Boday said, adding that the old model of individual vendors creating all the new features for an architecture no longer makes sense. “With today’s rate of change of technology, that is no longer a viable means for addressing those changes.”
IBM Power8 Server Comes With Nvidia’s NVLink Interconnect
Ian Buck, vice president of accelerated computing at Nvidia, said those changes are happening quickly.
“Data center workloads are changing,” Buck wrote in a post on the company blog. “Not long ago these systems were primarily used to handle storage and serve up web pages, but now they’re increasingly tasked with AI [artificial intelligence] workloads like understanding speech, text, images and video or analyzing big data for insights. Billions of consumers want instant answers to a multitude of questions, while enterprise companies want to analyze mountains of data to better serve their customers’ needs.”
GPUs from Nvidia and Advanced Micro Devices, as well as other accelerators, are being used by high-performance computing (HPC) organizations and increasingly by enterprises to improve the performance of their systems while controlling power consumption. Nvidia in April announced the massive P100 GPU, which uses the NVLink technology. The company is offering the P100 in its own system, the DGX-1, which officials called a supercomputer for AI and deep learning that includes eight of the GPUs and two Intel Xeon chips.
IBM with its latest Power8 server is the latest vendor to use the P100, but Nvidia officials have said they expect other systems makers—including Hewlett Packard Enterprise (HPE) and Dell—to roll out x86 systems with the GPU. The 2U (3.5-inch) Power S822LC runs on two Power8 chips with up to 20 cores and can hold up to four Nvidia P100 GPUs. Along with NVLink, the system also offers PCIe 3.0 and CAPI interconnect capabilities. It comes with up to 1TB of memory and can support hard drives or solid-state drives (SSDs) for storage.
Boday said the new server can cost as much as 30 percent less in some configurations when compared with x86-based servers. Pricing begins at $5,999.
The announcement of the S822LC comes a week after IBM officials gave more details about the company’s upcoming Power9 processors, which will offer up to 24 processing cores and be able to run an array of accelerators, including GPUs, FPGAs and application-specific integrated circuits (ASICs). The architecture also will embrace Nvidia’s upcoming NVLink 2.0 and PCI Express 4.0.