The Power Machine Learning and Deep Learning Reference Architecture (Release 1)

look_w

Machine and Deep Learning applications are one of the most exciting innovations in information
technology in this decade. Enabling developers to “program with data”, machine and deep
learning enables applications to make sense of data by classifying information based on
exemplary training patterns for the application.
This technology can be leveraged by a broad set of applications, from safety systems to personal
assistants to enterprise systems. Increasingly, driver assist technologies rely on machine and deep
learning patterns to recognize objects in a rapidly changing environment, personal digital
assistant technology is learning to categorize and group e-mail text message, and other content
based on their context. In the enterprise, machine and deep learning applications can identify
high value sales opportunities, enable smart call center automation, detect and react to intrusion
or fraud, and suggest solutions to technical or business problems.
The Power Machine Learning and Deep Learning (MLDL) Reference Architecture was created
to give deep learning developers and data scientists a platform on which to more quickly and
easily develop new machine learning-based applications and/or analyze data. It allows MLDL
application developers to construct simple applications that can easily be applied to more
complex, high-scale MLDL environments.
The Power MLDL Reference Architecture integrates software (the MLDL distro) and hardware
(the MLDL reference hardware) components. The combination has been tested to improve ease
of set-up, scalability, and performance of some of the most advanced and popular deep learning
frameworks in the research community: Theano, Caffe, and Torch. More will follow soon.
The MLDL distro provides an integrated, easy to install environment of multiple deep learning
frameworks. It will also provide over time a stable set of interfaces for future versions of the
Power MLDL Reference Architecture. This will allow users to preserve their innovation
investments as new framework features and OpenPOWER hardware advancements emerge.
The MLDL distro leverages the standard Ubuntu system install process. Hence, MLDL software
components can be installed with either the standard system installers (directly or from an
Internet-connected MLDL distribution server), or from a downloadable media file for servers
without Internet connectivity.
Application developers can execute their Deep Learning algorithms either on an OpenPOWER
CPU or a GPU. NVIDIA GPUs are accessed using device drivers and libraries provided by
NVIDIA. The current release of the Power MLDL Reference Architecture uses Ubuntu
14.04LTS in conjunction with the user-installed NVIDIA CUDA 7.5 device drivers and libraries,
which include the cuDNN library for deep neural network acceleration.
The current MLDL reference hardware includes a modular POWER8™ server with up to two
NVIDIA K80 GPU cards, which enables up to four GPUs.
Ordering the MLDL Reference hardware
Order an IBM Power System S822LC (models 8335-GCA or 8335-GTA) with 2 NVIDIA K80
Tesla GPUs. The K80s are optional for model 8335-GCA and included for model 8335-GTA.
The Redbook link below will help to ensure you have the order the right memory for each model.
Use the following specifications to plan for your servers, including dimensions, electrical, power,
temperature, environment, and service clearances.
Here is a sample model 8335 system configuration. Contact IBM for further ordering help.
MODEL FC DESCRIPTION DEFAULT MIN MAX

OS & FIRMWARE	2147 Primary OS - Linux 1 1 1 EC16 Open Power Abstraction Layer (OPAL) 1 1 1
PROCESSOR	EP00 8W P8 3.325GHZ PROC 190W (TURISMO) 0 * * EP01 10W P8 2.926GHZ PROC 190W (TURISMO) 2 * *
MEMORY	EM50 16GB (4X4GB) IS RDIMMS (1.35V) 1333MHZ 4GBIT DDR3 DRAMS 8 * * EM51 32GB (4X4GB) IS RDIMMS (1.35V) 1333MHZ 4GBIT DDR3 DRAMS 0 * * EM52 64GB (4X4GB) IS RDIMMS (1.35V) 1333MHZ 4GBIT DDR3 DRAMS 0 * * EM53 128GB (4X4GB) IS RDIMMS (1.35V) 1333MHZ 4GBIT DDR3 DRAMS 0 * *
DISKS	ELD0 1TB 7200 RPM SATA HDD IN MODIFIED GEN3 CARRIER 2 0 2 ES6A 2TB 7200 RPM 512E SATA HDD IN MODIFIED GEN3 CARRIER 0 0 2
GPU	EC49 COMPUTE INTENSIVE ACCELERATOR, PCIE 3 X16 GPU 2 0 2
ADAPTER	EC3A 2-PORT 40GbE RoCE SFP+ PCIe 3.0 LP ADAPTER (TRAVIS-3EN LP) 0 0 2 EC3E 2-PORT 100Gb EDR IB PCIe x16 LP CAPABLE ADAPTER - Twinax 0 0 1 EC3T 1-PORT 100Gb EDR IB PCIe x16 LP CAPABLE ADAPTER - Twinax 1 0 1 EL4M PCIe2 LP 4-port 1GbE Adapter 1 0 3 EN0T QUAD E'NET (2X1 + 2X10 10Gb) PCIe Gen 2 X8/SHORT LP-Fiber 0 0 3 EN0V QUAD E'NET (2X1 + 2X10 10Gb) PCIe Gen 2 X8/SHORT LP-Twinax 0 0 3
POWER	EB2V AC POWER SUPPLY, 1300 WATT (200-240V) 2 2 2 6458 PWR CBL, DRWR TO IBM PDU, 14', 200-240V/10A, IEC320/C13, IEC320/C14 2 2 2
NETWORK CABLE	EB5A 3M EDR IB / 100GbE Optical Cable QSFP 1 0 * ** customer needs to some additional Cat5e Cables **
MICS	0983 ASSEMBLED IN COMPLIANCE WITH US TRADE AGREEMENT ACT INDICATOR (ORDERABLE IN THE US ONLY) 0 0 1
OS FIRMWARE	9450 UBUNTU LINUX PROCESSOR COUNTER SPECIFY 24 0 24 EC16 OPAL BAREMETAL 1 0 1

Note: You can customize your CPU, memory, network adapter, and cables leng