HP NVIDIA Tesla M2070Q, NVIDIA Tesla M2090, NVIDIA Tesla K10, NVIDIA Tesla K20, NVIDIA Tesla K20X Specification

HP supports, on select HP ProLiant servers, computational accelerator modules based on NVIDIA® Tesla™ Graphical Processing Unit
(GPU) technology.

The following Tesla Modules are available from HP, for use in certain HP SL-series servers.

NVIDIA Tesla M2070Q 2-Slot Passive Module
NVIDIA Tesla M2090 2-Slot Passive Module
NVIDIA Tesla K10 Dual GPU PCIe Module
NVIDIA Tesla K10 Rev B Dual GPU Module
NVIDIA Tesla K20 5 GB Module
NVIDIA Tesla K20X 6 GB Module

The NVIDIA Tesla M2070Q module can also be used in HP ProLiant WS460c workstation blades.

Based on NVIDIA's CUDA™ architecture, the Tesla Modules enable seamless integration of GPU computing with HP ProLiant servers for
high-performance computing and large data center, scale-out deployments. These Tesla Modules deliver all of the standard benefits
of GPU computing while enabling maximum reliability and tight integration with system monitoring and management tools such as HP
Insight Cluster Management Utility.

The Tesla M2070Q uses the NVIDIA Fermi GPU architecture that combines Tesla's high performance computing - found in the other
Tesla Modules - and the NVIDIA Quadro® professional-class visualization in the same GPU. The Tesla M2070Q is the ideal solution for
customers who want to deploy high performance computing in addition to advanced and remote visualization in the same datacenter.

The HP GPU Ecosystem includes HP Cluster Platform specification and qualification, HP-supported GPU-aware cluster software, and
also third-party GPU-aware cluster software for NVIDIA Tesla Modules on HP ProLiant Servers. In particular, the HP Insight Cluster
Management Utility (CMU) will monitor and display GPU health sensors such as temperature. CMU will also install and provision the GPU
drivers and the CUDA software. The HP HPC Linux Value Pack includes a GPU-enhanced version of IBM Platform LSF, with the capability
of scheduling jobs based on GPU requirements. This capability is also available for HP in other popular schedulers such as Altair PBS
Professional, and Adaptive Moab.

NVIDIA K10

NVIDIA K20, K20X

What's New

Support for the NVIDIA Tesla K10 Rev B Dual GPU Module

QuickSpecs

NVIDIA Tesla GPU Modules for HP ProLiant Servers

Overview

DA - 13743 North America — Version 16 — September 30, 2013

Page 1

NVIDIA Passive Tesla
Modules

NVIDIA Tesla M2070Q 6GB GPU Graphics Module

A0C39A

NVIDIA Tesla M2075 6 GB Computational Accelerator

A0R41A

NVIDIA Tesla M2090 6 GB Computational Accelerator

A0J99A

NOTE:

2-slot passively cooled Tesla modules with 6 GB memory.

NOTE:

Please see the HP ProLiant SL250s Gen8 or SL270s Gen8 server QuickSpecs for
Technical Specifications and additional information.

http://h18004.www1.hp.com/products/quickspecs/14232_na/14232_na.html
http://h18004.www1.hp.com/products/quickspecs/14405_na/14405_na.html

NVIDIA Tesla K10 Dual GPU PCIe Computational Accelerator

NOTE:

Please see the HP ProLiant SL250s Gen8 or SL270s Gen8 QuickSpecs for
Technical Specifications and additional information.

http://h18004.www1.hp.com/products/quickspecs/14232_na/14232_na.html
http://h18004.www1.hp.com/products/quickspecs/14405_na/14405_na.html

B3M66A

NVIDIA Tesla K10 Rev B Dual GPU PCIe Computational Accelerator

NOTE:

2-slot passively cooled pair of Tesla GPUs each with 4 GB memory.

NOTE:

Please see the HP ProLiant SL250s Gen8 server QuickSpecs for Technical
Specifications and additional information.

http://h18004.www1.hp.com/products/quickspecs/14232_na/14232_na.html
http://h18004.www1.hp.com/products/quickspecs/14405_na/14405_na.html

E5V47A

NVIDIA Tesla K20 5 GB Computational Accelerator

C7S14A

NVIDIA Tesla K20X 6 GB Computational Accelerator

C7S15A

NOTE:

2-slot passively cooled Tesla GPUs based on NVIDIA Kepler architecture.

NOTE:

Please see the HP ProLiant SL250s Gen8 server QuickSpecs for Technical
Specifications and additional information.

http://h18004.www1.hp.com/products/quickspecs/14232_na/14232_na.html
http://h18004.www1.hp.com/products/quickspecs/14405_na/14405_na.html

QuickSpecs

NVIDIA Tesla GPU Modules for HP ProLiant Servers

Models

DA - 13743 North America — Version 16 — September 30, 2013

Page 2

M2070Q, M2090, K10, K20 and K20X Modules

Performance of the M2090 Module

512 CUDA cores
655 Gigaflops of double-precision peak performance
1330 Gigaflops of single-precision peak performance
GDDR5 memory optimizes performance and reduces data transfers by keeping large data sets in 6 GB of local memory that is
attached directly to the GPU.
The NVIDIA Parallel DataCache™ accelerates algorithms such as physics solvers, ray-tracing, and sparse matrix multiplication
where data addresses are not known beforehand. This includes a configurable L1 cache per Streaming Multiprocessor block and
a unified L2 cache for all of the processor cores.
The NVIDIA GigaThread™ Engine maximizes the throughput by faster context switching that is 10X faster than the M1060
module, concurrent kernel execution, and improved thread block scheduling.
Asynchronous transfer turbo charges system performance by transferring data over the PCIe bus while the computing cores are
crunching other data. Even applications with heavy data-transfer requirements, such as seismic processing, can maximize the
computing efficiency by transferring data to local memory before it is needed.
The high speed PCIe Gen 2.0 data transfer maximizes bandwidth between the HP ProLiant server and the Tesla processors.

Performance of the K10 Module

3072 CUDA cores (1536 per GPU)
190 Gigaflops of double-precision peak performance (95 Gflops in each GPU)
4577 Gigaflops of single-precision peak performance (2288 Gigaflops in each GPU)
GDDR5 memory optimizes performance and reduces data transfers by keeping large data sets in 8 GB of local memory, 4 GB
attached directly to each GPU.
The NVIDIA Parallel DataCache™ accelerates algorithms such as physics solvers, ray-tracing, and sparse matrix multiplication
where data addresses are not known beforehand. This includes a configurable L1 cache per Streaming Multiprocessor block and
a unified L2 cache for all of the processor cores.
Asynchronous transfer turbo charges system performance by transferring data over the PCIe bus while the computing cores are
crunching other data. Even applications with heavy data-transfer requirements, such as seismic processing, can maximize the
computing efficiency by transferring data to local memory before it is needed.
The high speed PCIe Gen 3.0 data transfer maximizes bandwidth between the HP ProLiant server and the Tesla processors.

Performance of the K20 Module

2496 CUDA cores

1.17 Tflops of double-precision peak performance

3.52 Tflops of single-precision peak performance
GDDR5 memory optimizes performance and reduces data transfers by keeping large data sets in 5 GB of local memory that is
attached to the GPU
The NVIDIA Parallel DataCache™ accelerates algorithms such as physics solvers, ray-tracing, and sparse matrix multiplication
where data addresses are not known beforehand. This includes a configurable L1 cache per Streaming Multiprocessor block and
a unified L2 cache for all of the processor cores.
Asynchronous transfer turbo charges system performance by transferring data over the PCIe bus while the computing cores are
crunching other data. Even applications with heavy data-transfer requirements, such as seismic processing, can maximize the
computing efficiency by transferring data to local memory before it is needed.
Dynamic Parallelism capability that enables GPU threads to automatically spawn new threads.
Hyper-Q feature that enables multiple CPU cores to simultaneously utilize the CUDA cores on a single GPU.
The high speed PCIe Gen 2.0 data transfer maximizes bandwidth between the HP ProLiant server and the Tesla processors.

QuickSpecs

NVIDIA Tesla GPU Modules for HP ProLiant Servers

Standard Features

DA - 13743 North America — Version 16 — September 30, 2013

Page 3

Performance of the K20X Module

2688 CUDA cores

1.32 Tflops of double-precision peak performance

3.95 Tflops of single-precision peak performance
GDDR5 memory optimizes performance and reduces data transfers by keeping large data sets in 6 GB of local memory that is
attached to the GPU
The NVIDIA Parallel DataCache™ accelerates algorithms such as physics solvers, ray-tracing, and sparse matrix multiplication
where data addresses are not known beforehand. This includes a configurable L1 cache per Streaming Multiprocessor block and
a unified L2 cache for all of the processor cores.
Asynchronous transfer turbo charges system performance by transferring data over the PCIe bus while the computing cores are
crunching other data. Even applications with heavy data-transfer requirements, such as seismic processing, can maximize the
computing efficiency by transferring data to local memory before it is needed.
Dynamic Parallelism capability that enables GPU threads to automatically spawn new threads.
Hyper-Q feature that enables multiple CPU cores to simultaneously utilize the CUDA cores on a single GPU.
The high speed PCIe Gen 2.0 data transfer maximizes bandwidth between the HP ProLiant server and the Tesla processors.

Reliability

"ECC Memory meets a critical requirement for computing accuracy and reliability for datacenters and supercomputing centers. It
offers protection of data in memory to enhance data integrity and reliability for applications. For M2075, M2070Q, M2090, K20,
K20X register files, L1/L2 caches, shared memory, and DRAM all are ECC protected. For K10, only external DRAM is ECC
protected. Double-bit errors are detected and can trigger alerts with the HP Cluster Management Utility. Also, the Platform LSF
job scheduler, available as part of HP HPC Linux Value Pack, can be configured to report when jobs encounter double-bit errors.
Passive heatsink design eliminates moving parts and cables reduces mean time between failures.

Programming and Management Ecosystem

The CUDA programming environment has broad support of programming languages and APIs. Choose C, C++, OpenCL,
DirectCompute, or Fortran to express application parallelism and take advantage of the innovative Tesla architectures. The
CUDA software, as well as the GPU drivers, can be automatically installed on HP ProLiant servers, by HP Insight Cluster
Management Utility.
"Exclusive mode" enables application-exclusive access to a particular GPU. CUDA environment variables enable cluster
management software such as the Platform LSF job scheduler (available as part of HP HPC Linux Value Pack) to limit the Tesla
GPUs an application can use.
With HP ProLiant servers, application programmers can control the mapping between processes running on individual cores, and
the GPUs with which those processes communicate. By judicious mappings, the GPU bandwidth, and thus overall performance,
can be optimized. The technique is described in a white paper available to HP customers at:

www.hp.com/go/hpc

. A heuristic
version of this affinity-mapping has also been implemented by HP as an option to the mpirun command as used for example
with HP-MPI, available as part of HP HPC Linux Value Pack.
GPU control is available through the nvidia-smi tool which lets you control compute-mode (e.g. exclusive),
enable/disable/report ECC and check/reset double-bit error count. IPMI and iLO gather data such as GPU temperature. HP Cluster
Management Utility has incorporated these sensors into its monitoring features so that cluster-wide GPU data can be presented
in real time, can be stored for historical analysis and can be easily used to set up management alerts.

QuickSpecs

NVIDIA Tesla GPU Modules for HP ProLiant Servers

Standard Features

DA - 13743 North America — Version 16 — September 30, 2013

Page 4