GPU - InDesign
SUPERMICR®
SAINS ES
Shattering the 1U Server Performance Record
UPERMICRO RECENTLY ANNOUNCED a new class of serv-
ers that combines massively parallel GPUs with multi-core CPUs
in a single server system. This unique configuration delivers per-
formance at least an order of magnitude better than traditional
quad-core CPU-based servers. This breakthrough technology im-
mediately provides users with the ability to implement tasks that
were traditionally addressed only with massive supercomputers
or that were simply unsolvable. Supermicro calls this new server
class, the latest in its rich history of technology innovations, the
GPU Supercomputing Server.
Storage
High-Density SASISATA
Market Opportunity a
+ 25-2708
+ GPU Optimized
FIRST TO MARKET
Supermicro demonstrated its first GPU Supercomputing Serv-
ers, the SuperServer 6016 TGF series, at the Computex 2009 Show
in Tai ly June. This server seri {ual Intel® Xeon®
Processor 5500 series (Nehalem) and two PCI-Express 2.0 x16 in-
terfaces to support multiple GPUs. With two double-width GPUs
and double PCI-Express 2.0 x16 lanes, this 1U server delivers truly
non-blocking GPU performance, or up to 2 Teraflops of process-
ing power, which makes it the fastest 1U server on the planet
‘The SuperServer 6016 TGF 1U series is the first of an entire line
SuperBlade®
- OfficeBlade™ (508)
=» DatacenterBlade™
U Twin m
Whisper-Quiet Award-winning Ann
Workstations X
GPU Supercomputing Server
Ultra-high performance
Scientifically optimized
Chassis & Power
Supply = шо 6 (Universal vo) A
- fay = ei Ce Best Performance per Wait
(375GFLOPSIkW]
FE dE pr day (TH,
ET
se Redundant 1U
Rackmount Rackinalin
Solutions
1998 2002 2005 2007 2008 2009
Supermicro Server Building Block Solutions® Technology Progression
Figure 1: Supermicro Product and Market Opportunity Growth
Systems in Design - July 2009
SHATTERING THE 1U SERVER PERFORMANCE RECORD
Figure 2: Supermicro GPU Supercomputing Servers.
Left to right: SS6016GT-TF-TM2, SW7046GT-TRF
of GPU-optimized systems Supermicro has created to meet the
requirements of the high-performance computing market. By
the end of June, Supermicro will launch a 4U/Tower system, the
SuperWorkstation 7046 T-GRE which supports four double-width
GPUs. These platforms feature Supermicro new Gold Level!
(93% efficiency) power subsystems to deliver breakthrough per-
formance-per-watt.
WHAT IS A GPU SUPERCOMPUTING SERVER?
To understand GPU Supercomputing Servers, we first need
to understand the architecture of a GPU. GPUs (Graphics
Processing Units) are high-performance multi-core proces-
sors capable of very high computation and data throughput.
Once specially designed for computer graphics and difficult
lo program, todays GPUs are general-purpose parallel pro-
cessors that support standard application programming in-
terfaces (e.g. APIs like OpenGL) and industry-standard lan-
guages such as C. Applications that are run on GPUs often
achieve substantial speed increases compared with optimized
CPU-only implementations.
The model for GPU supercomputing is to use a CPU and
GPU together in a heterogeneous (e.g. mixed) supercomput-
ing model. The sequential part of the application runs on
the CPU and the computationally-intensive part runs on the
GPU. From the user's perspective, the application just runs
faster because it uses the high performance of the GPU to
boost computing speed, This operation is sometimes called
GPGPU (General-Purpose computation on Graphics Pro-
cessing Units).
Figure 3: Heterogeneous Computing Model. Left to right: CPU, GPU
El Systems in Design - July 2009
Such leaps in performance do not come eas-
ily. Application developers must modify their
code by taking the compule-intensive por-
tions and mapping them to the GPU. The rest
of the application remains on the CPU. Map-
ping à function to the GPU involves rewriting
the functions to expose their parallelisms and
adding the proper “C” programming language
instructions to move data to and from the
GPU efficiently. New parallel processing archi-
tecture tools such as CUDA help to automate
this process, saving programming resources
and expanding the applications universe for
CPU+GPU heterogeneous supercomputing.
A GPU Supercomputing Server integrates GPUs into the
server architecture. The Supermicro S56016GT-TF-TM2, Гог
example, utilizes two NVIDIA Tesla M1060 double-width GPU
cards, each of which has 240 processor cores generating up to
1 Teraflop, for a total computing capability of up to 2 Teraflops.
This powerful system, with intelligent cooling control technol-
ogy, is optimized for mission critical server cluster applica-
tions. The 4U/Tower Supermicro SW7046GT-TRF integrates
four double-width GPUs for a total of up to 4 Teraflops.
WIDEST VARIETY OF GPU SUPERCOMPUTING SERVERS
Supermicro has developed the widest variety of GPU Su-
percomputing Servers in the industry. These include the 55-
6016GT series of four 1U servers announced at Computex and
two 4U/Tower systems, the SW7046GT-TRF and SW7046A-
HR+, These systems are identified in bold in Table 1 below to-
gether with the maximum number of double-width GPUs and
expansion slots that they can support simultaneously.
Please review the Supermicro Supercomputing Server Solu-
tions web page for more information:
http://www.supermicro.com/GPU/
Supermicros GPU Supercomputing Server family also in-
cludes a growing number of systems that can be customized
using Supermicros unique Server Building Block Solutions”.
These systems are identified in Table 1 by the number of GPUs
and expansion slots that they can support. Initially four 1U and
4U/ Tower chassis are supported and a wide variety of Super-
micros next-generation motherboards supporting the Intel®
Xeon® Processor 5500 Series (Nehalem) are also available as
Building Blocks.
More information on these Supermicro chassis
at: htlp://www.supermicro.com/products/chassis/
Additional motherboard information can be found at:
ИХеоп1333/
is available
r tel
1 [y ir я ; du
http: www.supermicro.com/ products,
! 80Plus.org
MotherBoard
Table 1: Supermicro GPU Supercomputing Servers (Bold) and Server Building Block Solutions™
SHATTERING THE 1U SERVER PERFORMANCE RECORD
SCT43T0-865B-50 (40/TOWER)
X6DAH-+ SWT046A-HR+
3 GPUs, 1 Expansion Slot
2 GPUs*, 3 Expansion Slot 3 GPUs*, 1 Expansion Slot
2 GPUs*, 3 Expansion Slot 3 GFUs*, 1 Expansion Slot
SW7046A-3: 2 GPUs, 7 Expansion Slots — 2 GPUs, 2 Expansion Slots
SWI046A-T: 7 GPUs, 7 Expansion Slots 2 GPUs, 2 Expansion Slots
X8DTH-6/6F
LETT RT
LIT Kl
X&DTT-IBOF
SCT4570-R1200B (4UTOWER)
Chassis
58186-14008 (1U RACKMOUNT)
SCT47TO-R14008 (4U/TOWER)
SWT046T-
4 GPUs, 3
GRF
Expansion Slots
SS60166T-TF-TM2; Enterprise Level
2 integrated GPUs, 1 Expansion slot {low Profile)
SS6016ET-TF-TC2: 2 integrated GPUs, 1 Expansion
slot (low Profile)
SS6016ET-TF: 2 GPUs, 1 Expansion slot (low Profile)
SS6016XT-TF: 4 Expansion slots + 1 Expansion slot
{low profile)
3 GPUs*, 1 Expansion Slot
3 GPUs*, 1 Expansion Slot
2 GPUs, 2 Expansion Slots
2 GPUs, 2 Expansion Slots
SS6016T-6TF: 1 GPU
SS6016T-GIBXF: 1 GPU
S56016T-6IB0F: 1 GPU
ADVANTAGES OF SUPERMICRO GPU
SUPERCOMPUTING SERVERS
Superimicro’s GPU Supercomputing Server family provides
many feature advantages of interest to the user community.
Widest Selection
Supermicro’ wide line of GPU systems, outlined in Table 1,
can satisfy any GPU Supercomputing server application. The
line comprises a wide variety of 1U and 4U/ Tower systems,
chassis and motherboards.
Highly Reliable Thermal Optimization
Supermicro advanced thermal subsystem designs opti-
mize the cooling elements of the server with the thermal char-
acteristics of the GPU, CPU, motherboard, and other compo-
nts. The cooling system is designed to protect against any
single point of fan failure, enhancing system reliability and
simplifying maintenance. In addition, by monitoring both
CPU and GPU thermal readings through the I*C bus, Super-
micros unique intelligent cooling control and auto fan-speed
adjustment keep the entire system operating smoothly.
Direct Connect Architecture
The Supermicro design architecture provides direct PCI-
Express 2.0 x16 non-blocking connectivity to each GPU to
take advantage of the GPU's full bandwidth. No extra cabling
is required, since the GPUs are connected directly to the server
motherboard via riser cards, thus improving reliability, airflow,
and maintenance, as well as reducing costs.
El Systems in Design - July 2009
*Some Gen2 PCLE x8 in x16 slot
** Requires video output card
Industry-Leading Power Efficiency
Supermicros GPU Supercomputing server platforms fea-
ture Gold Level (93% efficiency) power subsystems, along
with high-efficiency motherboard and thermal designs to
deliver breakthrough performance-per-watl and increased
system reliability.
Flexible Networking Connectivity
The §S6016GT GPU Supercomputing Server series in-
cludes a single half-height PCI-E 2.0 slot for additional con-
nectivity. The SW7046GT-TRF 4U/ Tower server with 4 dou-
ble-width GPUs has 3 additional PCI-E expansion slots for
add-on cards.
Advanced System Management
Supermicro’s advanced IPMI capability allows the GPU sys-
tem to be managed directly via remote monitor and controlling
functions. All the key elements of the server system- CPU, GP
and power supply- are remotely monitored from Supermicro’s
extensive onboard IPMI management and control subsystem.
APPLICATIONS AND INDUSTRIES
This new line of highly parallel, many-core, multi-GPU sys-
tems is an excellent choice for an extensive range of graphics
and computationally intensive applications in a wide variety of
industries. In general, these systems are expected to make the
fastest teraflop clusters much more affordable and accessible
for users throughout the world.
Applications requiring high arithmetic intensity, such as
SHATTERING THE 1U SERVER PERFORMANCE RECORD
Quantum Chemistry
Medical Imaging Oil & Gas Exploration
Figure 4: GPU Supercom
dense linear algebra, partial differential equations, n-body
problems, and finite difference formulas, can be easily ac-
commodated using Supermicro Supercomputing Servers,
High bandwidth problems such as sequencing (virus scan-
ning, genomics), sorting, and database problems are also
attractive applications. Finally, visual computing problems
such as graphics, image processing, lomography, and ma-
chine vision, being original applications for GPUs, are espe-
cially accessible.
Fields where these applications are common include med-
ical imaging, oil and gas exploration, computational chem-
istry, telecommunications, financial simulation, weather
forecasting, and astrophysics. In addition research and engi-
neering projects and general scientific computing investiga-
tions can also benefit from Supermicro’s GPU Supercom-
puting Servers, BOE
Financial Simulation 3D Ultrasound Astrophysics
puting Server Markets
ADDITIONAL RESOURCES
Those wishing to learn more about this topic may find the
following resources helpful places to start:
Supermicro GPU
hUp://wwwsupermicro.com/GPU/
GPGPUorgisa central resource for GPGPU news and information:
hitp://gpgpu.org/
Dr. Dobb's Journal discusses applications software for GPUs:
http://www.ddj.com/hpc-high-performance-comput-
ing/206900471
ATI Stream Technology Site:
http://ati.amd.com/technology/streamcomputing
NVIDIAs CUDA Zone:
http://www.nvidia.comfobject/cuda_home.html
Who's working on GPU apps: “GPGPU People”
http:/fwww.gpgpu.org/w/index,php/GPGPU_People
Was this manual useful for you? yes no
Thank you for your participation!

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project

Download PDF

advertising