BLOG: Make the Most of Your AI Models with the New IBM Power System Data/Cloud/Inferencing Server

March 30th, 2020 BLOG: Make the Most of Your AI Models with the New IBM Power System Data/Cloud/Inferencing Server

Ron Gordon
Director » Power Systems

In late January, IBM Power Systems announced a new server developed for data, cloud, and inferencing applications. The server is named the IBM Power System IC922, which stands for Inferencing Cloud, POWER9, 2 socket, 2U, and is now available in the general availability (GA) 1.0 configuration.

The IBM Power System IC922 is a Linux-only server and is a logical replacement and improvement over the IBM Power Systems LC921 and LC922, servers that use the current IBM POWER9 processor-based technology and support Linux operating systems.

A high-level comparison of these systems demonstrates several distinct advantages of the IC922 over the LC922 and the LC921 (for denser cluster requirements):

  1. Speed The new IC922 has 24, 32, or 40 core variants with speeds up to 3.35 GHz, which provides higher core performance than the 2.9 GHz LC922 system.
  2. Component Connection The IC922 has 10 PCIe slots with support for PCIe Gen4 versus six PCIe slots for the LC922.
  3. Storage The internal storage of the IC922 supports 24 bays while the LC922 has 12 bays.
  4. Memory Bandwidth The memory of the IC922 is 2x that of the LC922, 340 GB/s vs. 170 GB/s, and it can handle up to 2 TB while the LC922 only supports up to 1 TB.
  5. Security There are added hardware security features in the POWER9 processor of the IC922.
  6. IBM-Made The IBM Power System IC922 is manufactured by IBM, whereas the LC922 is manufactured by a non-IBM company.

Customers looking at the IBM Power System LC922 for Linux clusters should now strongly consider the IBM Power System IC922 to support data, cloud, and artificial intelligence (AI) inferencing applications provisioned in both clusters and standalone architectures.

Nvidia T4 GPU vs. Nvidia V100 GPU

The Power System IC922 can support up to six Nvidia T4 GPUs in its PCIe slots, however the Nvidia T4 is a lower power and lower performance GPU than the Nvidia V100 GPU found in the IBM Power System AC922. The Nvidia T4 GPU has half the number of Nvidia Tensor GPU cores than the V100 GPU (320 vs. 640) and, generally speaking, provides half the teraflops (8.1 TFLOPS vs. 16 TFLOPS)—meaning the Nividia T4 processor can calculate half the number of floating-point operations per second.

The Nvidia V100 GPU on the IBM Power System AC922 is directly connected to the P9 processor via a proprietary NVLink supporting 300 GB/s. This difference indicates that for AI applications, the IBM Power System AC922 is better equipped to handle model training while the IBM Power System IC922 should be used for data preparation and model inferencing. The same optimized software (Watson Machine Learning Community Edition, Watson Machine Learning Accelerator, Watson Visual Insights, IBM Cloud Pak for Data) is supported on both the AC922 and the IC922.

The Nvidia T4 GPUs are optional on the IBM Power System IC922 and should mainly be configured when running AI models for prediction on live data in real-time, called inferencing or scoring. Without the GPUs, the IC922 is a low-cost, high-performance, highly reliable server supporting the data and cloud markets. In the data and cloud segment, the support of the family of IBM Cloud Paks provides the software infrastructure needed to optimize these solutions.

The above information applies to the GA 1.0 version of the IBM Power System IC922. In an IBM Statement of Direction, what IBM calls the GA 1.5 version will become available in 2020 (deferred due to testing requirements). GA 1.5 will support up to eight Nvidia T4 GPUs, Red Hat Enterprise Linux 8.1 and Red Hat Enterprise Virtualization (RHEV), and include the addition of NVMe drives in the storage bays. GA 1.5 will be able to be applied as an upgrade to the current GA 1.0 version. After the GA 1.5 release, IBM will release the GA 2.0 version, which will introduce new FPGA and ASIC adapters for high-performance acceleration in IO and compute. (GA 1.5 and GA 2.0 are Statements of Direction by IBM and dates and content are subject to change).

That was a lot of data to absorb—but if you’re looking for even more, the IBM Redbook for the IC922 can be downloaded here.

For more information on how IBM Power Systems can support your business, please contact your Mainline representative or click here to contact us with any questions..

Related Information:

IBM Power Systems Software Announcements for 4th Quarter 2019

Submit a Comment

Your email address will not be published. Required fields are marked *