IBM unveiled its latest microprocessor Tuesday in conjunction with a new server line that better integrates the new CPU with other embedded devices to power advanced workloads like artificial intelligence.
The new AC922 Servers, along with the Power9 processors, deliver unprecedented performance for computationally intensive tasks like training machine-learning models and advanced analytics, Sumit Gupta, vice president for high performance computing and data analytics at IBM, told CRN.
Those servers directly attack the problem that's slowing the advent of artificial intelligence -- slow connectivity in the system bus that limits data flow between various chips on the server board.
"The server is going from being just about CPU to becoming a rich server with lots of different devices," Gupta said, such as graphics and network accelerators that offload computation from the central processor.
The AC922s realize the vision of the OpenPower Foundation that IBM launched roughly five years ago, Gupta said.
Back then, IBM had the realization "that the CPU was no longer really delivering the performance needs required for modern workloads," he said, such as data analytics, machine learning and now deep learning.
IBM concluded speeding those tasks wasn't going to take faster microprocessors, but a faster internal bus connecting the various compute acceleration components, like graphics processing units.
"We knew these other devices were going to be critical in the server, networking devices, other processors. We needed to improve the data throughput, how data moved within a server," he said.
"That's the answer to Moore's Law slowing."
Those efforts are now bearing fruit, Gupta added, due to massive interest in recent years in artificial intelligence. GPU computing "only recently it hit its stride," he said.
Power9, according to Gupta, is the first PCI-Express chip with fourth-generation capabilities. Combined in the AC922 with Nvidia NVlink 2.0 interconnections and leveraging OpenCAPI architecture, it accelerates the flow of data through the server almost 10 times faster than chips enabled to run PCI-Express 3.0 (third-generation) connections, like the ones Intel offers on its most-advanced x86 processors, he said.
NVLink 2.0 is a propriety interface between only IBM Power units and Nvidia Volta GPUs.
That innovation doubles processing speed, he said, giving IBM "a significant data-movement performance advantage."
Deep learning workloads have been clocked to run roughly 3.5 times faster on the new servers, Gupta said.
AC922s will start shipping later in December, with the first batch going to the Oak Ridge National Laboratory in Tennessee, where thousands will be clustered to build the world's fastest supercomputer, Gupta said.
In addition to the high-performance supercomputing market, enterprises running AI applications, especially deep learning, represent another base of buyers for the new servers, Gupta said.
Google will also be a customer, the internet giant said in a statement.
And while there are no announcements on the cloud front yet, "nearly all of these IBM technologies eventually make it to IBM cloud," he said.