Systems

Data Sheet

IBM Power Advanced

Compute (AC) AC922
Server
The Best Server for Enterprise AI

Modern AI, HPC and analytics workloads are driving an ever-growing

Highlights set of data intensive challenges that can only be met with accelerated
infrastructure.
IBM® Power Systems™ Accelerated
Compute (AC922) server is an accelera-
tion superhighway to enterprise-class AI.
To help meet these demands, IBM Power Systems has designed AC922
for the AI Era, the best server for Enterprise AI. The AC922 leverages
A Superhighway to and for acceleration—
IBM’s new POWER9™ processor with a myriad of modern connectivity
●● ● ●

unleashes more accelerated computing

potential, in the post CPU-only era capabilities yielding up to 5.6x1 the data movement over the antiquated
PCIe Gen 3 buses found in x86. IBM Power Systems deliver the only
●● ● ●
Designed for the AI Era, architected for
the modern analytics and AI workloads architecture enabling NVLink between CPUs and GPUs, unlocking new
that fuel insights potential for accelerated computing.
Delivering enterprise-class AI - cutting-
●● ● ●

edge AI innovation data scientists desire,

with dependability IT requires

●● ● ●
Proven deployments from small clusters
to the world’s largest supercomputers,
with near linear scaling

Simplest GPU acceleration3 —focus

●● ● ●

on what to accelerate not how, with

coherence, available with AC922 in
addition to larger model sizes enabled

Power System AC922 internal components

Systems
Data Sheet

IBM Power Systems Accelerated Compute

(AC922) Servers
IBM Power Systems Accelerated Compute (AC922) server is an
IB

acceleration superhighway to enterprise-class AI. PCIe Gen 4

CAPI 2.0

170GB/s
The next generation of IBM Power Systems, with POWER9 CPU

technology, is built with innovations that deliver unprecedented DDR4

speed for the AI Era. The IBM Power System AC922 for high NVLink
100GB/s
NVLink
100GB/s
performance computing provides: Coherent access
to system memory
(2TB)
NVLink
●● ●
Faster I/O -The AC922 includes a variety of next-generation 100GB/s

I/O architectures, including: PCIe gen4, CAPI 2.0,

OpenCAPI and NVLINK. These interconnects provide up NVLink NVLink
to 5.6 times1 as much bandwidth for today’s data-intensive 100GB/s 100GB/s
NVDIA V100 NVDIA V100 NVDIA V100
workloads versus the antiquated PCIe gen3 found in
x86 servers.
●● ●
Extraordinary CPUs - While blazingly fast on their own,
Power System AC922 with 6 GPUs and water Cooling
POWER9 CPUs truly excel in their ability to unleash and
exploit the performance of everything around them. Built
for the AI-Era, the POWER9 supports up to 5.6x1 more ●● ●
Simplest AI architecture3 - AI models grow large, easily
I/O and 2x more threads than its x86 contemporaries. outgrowing GPU memory capacity in most x86 servers.
The POWER9 is available on configurations with 16, 18, CPU to GPU coherence in the AC922 addresses these con-
20 and 22 cores, for up to 44 cores in the AC922 server. cerns by allowing accelerated applications to leverage System
memory as GPU memory. And it simplifies programming by
●● ●
Advanced GPUs -The AC922 pairs what is arguably the best
eliminating data movement and locality requirements. And,
GPUs for Enterprise AI with the best platform f or them.
by leveraging the 5.6x1 faster NVLink interconnect, sharing
The AC922 pairs POWER9 CPUs and NVIDIA Tesla V100
memory between CPUs and GPUs doesn’t bottleneck down
with NVLink GPUs. Delivers up to 5.6x1 times the perfor-
to PCIe 3 speeds, as it would on x86 servers.
mance for each pairing. This is the only server capable of
delivering this I/O performance between CPUs and GPUs. ●● ●
Enterprise-ready - Simplifies deep-learning deployment and
This provides massive throughput capability for HPC, performance. Unlocks a new, simpler end-to-end toolchain
deep learning and AI workloads. for AI users. Proven AI performance and scalability enable
you to start with one node, then scale to a rack or thousands
●● ●
1st PCIe generation 4 server - The AC922 is the industry’s
of nodes with near linear scaling efficiency.
first server to feature the next generation of the industry
standard PCIe interconnect. PCIe generation 4 delivers ●● ●
Built for the world’s biggest AI challenges, ideal for
approximately 2x the data bandwidth2 of the PCIe yours - The AC922 is the backbone of the CORAL Summit
generation 3 interconnect found in x86 servers. supercomputer, meeting milestones to deliver 200+ petaf lops
of HPC and 3 exaflops of AI as a service performance. But
with its efficiency and ease of AI deployment, it’s also ideally-
suited to address your organization’s AI aspirations.

2
Systems
Data Sheet

Designed for leaders in AI and Deep Learning, HPC and

high-performance analytics, the IBM Power System AC922
IB provides:
PCIe Gen 4
CAPI 2.0 ●● ●
2x POWER9 with NVLink 2.0 CPUs, with 16 DIMM
170GB/s sockets and up to 2 TB of memory
CPU
●● ●
A differentiated platform for GPU acceleration
– POWER9 with NVLink 2.0 Technology: a link with
DDR4

NVLink NVLink
Coherent access
150GB/s 150GB/s
up to 5.6X the performance (100 GB/sec air cooled or
to system memory
(2TB) 150 GB/sec water cooled) to each NVIDIA V100 with
NVLink GPU
– Incredible CPU to GPU and GPU to GPU
communication: up to 5.6X the data flow
NVLink
150GB/s (100 or 150 GB/sec) between adjacent NVIDIA®
NVDIA V100 NVDIA V100
Tesla® V100 GPU Accelerators on the same socket of
PCI-E Gen3 x16 solutions
●● ●
Easy GPU Programming—full coherence and access to
Power System AC922 with 4 GPUs and Air Cooling systems memory
●● ●
Advanced Mellanox ConnectX-5 InfiniBand Fabric, with
industry-leading PCIe Gen4 interface
●● ●
Optional CAPI-attached NVMe storage for exceptionally fast
storage I/O and burst buffer data staging
●● ●
Advanced interfaces to all accelerators: NVLink,
OpenCAPI™, CAPI 2.0 and PCIe Gen4

3
Systems
Data Sheet

Proven AI performance and scalability—

start with one node, then scale to a rack
much higher with near linear scaling
efficiency.

Power System AC922 POWER9 CPU

Power System AC922 internal view

Power System AC922 front view

4
Systems
Data Sheet

Power System AC922 (8335-GTC, 8335-GTW) at a glance

System configurations

Microprocessors 2x POWER9 with NVLink CPUs 16, 20 cores, or 18, 22 cores (with liquid cooling)

Level 2 (L2) cache 512 K

Level 3 (L3) cache 10 MB

RAM (memory) Up to 2 TB, from 16 DDR4 RDIMM Sockets

Internal disk storage 2x SFF (2.5") drive bays, optional NVMe SSD support in PCIe slots

Processor-to-memory 170 GB/s per socket, 340 GB/s per system

bandwidth

L2 to L3 cache bandwidth 7 TB/s on chip bandwidth

Internal SCSI disk bays n/a

Media bays n/a

Adapter slots 4 or 6 SXM 2.0 sockets, for NVIDIA Tesla V100 GPU Accelerators with NVLink.
2x PCIe x16 4.0 slots
1x PCIe16x ( x8,x8) 4.0 slot (multi-socket host direct supported)
1x PCIe x4 4.0 slot

Standard features

I/O ports 2x USB 3.0, 2x 1 GB Eth, VGA

Connectivity support (optional)

POWER Hypervisor™ KVM

Advanced POWER
Virtualization1 (option)

RAS features Processor instruction retry

Selective dynamic firmware updates
Chip kill memory
ECC L2 cache, L3 cache
Service processor with fault monitoring
Hot-swappable disk bays
Redundant cooling fans

Operating systems Red Hat Enterprise Linux, Ubuntu Linux

Power requirements 200 V to 240 V

System dimensions Width: 441.5 mm (17.4 in.)

Depth: 822 mm (32.4 in.)
Height: 86 mm (3.4 in.)
Weight: 30 kg (65 lbs.)

Warranty 3-year limited warranty, CRU (customer replaceable unit) for all other units (varies by country)
next business day 9am to 5pm (excluding holidays), warranty service upgrades and maintenance are available.

5
Why IBM?
IBM is a trailblazer in AI—From early machine learning
system in IBM Research to Watson® on Jeopardy, AI isn’t just
a buzzword for IBM. And we’re applying that innovation to
cognitive infrastructure, helping our customers on their
journey to AI. © Copyright IBM Corporation 2017

IBM Systems
IBM aligns cutting-edge innovation with enterprise New Orchard Road
Armonk, NY 10504
dependability—IBM has over 105 years of aligning continuous
innovation with our customers’ business needs. Produced in the United States of America
December, 2017

IBM is your partner for the AI Era—IBM provides the most IBM, the IBM logo, ibm.com, Power Systems, and POWER, are
trademarks of International Business Machines Corp., registered in many
f lexible and comprehensive range of technology and services jurisdictions worldwide. Other product and service names might be
needed for your entire journey to AI, whether you’re looking to trademarks of IBM or other companies. A current list of IBM trademarks
deploy a few nodes in order to investigate AI, or you’re looking is available on the web at “Copyright and trademark information” at
ibm.com/legal/copytrade.shtml
­ ­
to deploy exaflops of AI as a Service.
Linux is a registered trademark of Linus Torvalds in the United States,
other countries, or both.
For more information
To learn more about the Power System AC922 NVIDIA, NVIDIA Volta, NVIDIA NVLink are trademarks of NVIDIA
Corporation in the United States, other countries, or both.
please contact your IBM representative or
IBM Business Partner, or visit the following website: This document is current as of the initial date of publication and may be
changed by IBM at any time. Not all offerings are available in every
ibm.com/marketplace/ai-hpc-server-power-ac922 country in which IBM operates.

The performance data and client examples cited are presented for
Additionally, IBM Global Financing provides numerous pay-
illustrative purposes only. Actual performance results may vary depending
ment options to help you acquire the technology you need to on specific configurations and operating conditions.
grow your business. We provide full lifecycle management of
THE INFORMATION IN THIS DOCUMENT IS PROVIDED
IT products and services, from acquisition to disposition. “AS IS” WITHOUT ANY WARRANTY, EXPRESS OR
For more information, visit: ibm.com/financing
­ IMPLIED, INCLUDING WITHOUT ANY WARRANTIES
OF MERCHANTABILITY, FITNESS FOR A PARTICULAR
PURPOSE AND ANY WARRANTY OR CONDITION OF
NON-INFRINGEMENT. IBM products are warranted according to the
terms and conditions of the agreements under which they are provided.

Actual available storage capacity may be reported for both uncompressed

and compressed data and will vary and may be less than stated.
1 5.6x more I/O bandwidth – tested results are based on IBM Internal
Measurements running the CUDA H2D Bandwidth Test Hardware:
Power AC922; 32 cores (2 x 16c chips), POWER9 with NVLink 2.0;
2.25 GHz, 1024 GB memory, 4xTesla V100 GPU; Ubuntu 16.04. S822LC
for HPC; 20 cores (2 x 10c chips), POWER8 with NVLink; 2.86 GHz,
512 GB memory, Tesla P100 GPU Competitive HW: 2x Xeon E5-2640 v4;
20 cores (2 x 10c chips) / 40 threads; Intel Xeon E5-2640 v4; 2.4 GHz;
1024 GB memory, 4xTesla V100 GPU, Ubuntu 16.04
2 PCIe Generation 4 provides 2x data throughput vs. PCIe gen 3 (31.5 GB/s
vs 15.8 GB/s x16)
3 Simplest Ai Architecture - Coherence simplifies coding by abstracting
data movement and locality for developers

Please Recycle

POD03143-USEN-00
­ ­