IBM's new Power8 server packs in Nvidia's speedy NVLink interconnect

Nvidia's NVLink interface in IBM's new server gives a five-fold speed boost to communications between CPUs and GPUs

IBM is making headlines with its quantum computing research and brain-like chip called TrueNorth, but it also is bringing interesting technologies to its current Power server lineup.

Inside IBM's new S822LC server for high-performance computing is a new interconnect that gives a five-fold speed boost to communication between a CPU and graphics processor.

The interconnect is based on Nvidia's homegrown NVLink technology, which has been in the works for years. IBM's two-socket server, which is based on Power8 CPUs, is among the first available with the interconnect.

NVLink is essentially an upgrade to PCI-Express 3.0, which has been used for communication between a GPU and other components in a system.

Nvidia has built GPUs based on its new Pascal architecture that support NVLink, while IBM has added support for the NVLink I/O bus, called NVBus, in its chip.

IBM is among the first hardware vendors to ship an NVLink-compatible server. Nvidia has said x86 servers from companies like Dell, Hewlett Packard Enterprise, and Cray will ship next year.

Nvidia is already shipping a homegrown server called DGX-1 -- which has eight Tesla P100 GPUs -- with the NVLink interface. It is priced at a whopping $129,000.

Servers are using GPUs as co-processors to speed up applications, and a faster pipe to the CPU speeds up processing. Faster processing helps run databases, as well as fraud detection and engineering-related applications, said Dylan Boday, senior offering manager for Linux on Power infrastructures at IBM.

A good example of an application that will benefit is Kinetica, an in-memory database application that harnesses the computing power of GPUs. The application is designed to work with the NVLink interconnect, Boday said.

For now, only the Tesla P100 GPU will take advantage of the screaming NVLink data transfer speeds, and up to four GPUs can be plugged in the 2U server. The Linux server also has a PCI-Express 3.0 interconnect for other GPUs.

The two-socket Power server supports up to 20 CPU cores and 1TB of memory, and it has storage slots for hard drives and SSDs. It also has three PCI-Express slots. It has another interface called CAPI (coherent accelerator processor interface) for attaching custom chips like FPGAs (field programmable gate arrays) and custom chips.

IBM didn't provide a price for the server. The server will ship worldwide this month.

The Power8 server has the first version of NVLink. IBM plans to release Power9 chips and servers next year with support for NVLink 2.0, which will be significantly faster.

IBM on Thursday also announced the two-socket S822LC for Big Data server, which doesn't have the NVLink interface, but supports up to Nvidia K80 GPUs based on the older Kepler architecture. It supports up to 20 CPU cores, up to 512GB of memory, and 96TB storage in multiple slots. It also has five PCI-Express 3.0 and four CAPI slots.

Copyright © 2016 IDG Communications, Inc.