PNY TESLA K40 User Manual [ru]

NVIDIA® TESLA® K40 Module

NVIDIA® TESLA® K40

WORLD’S FASTEST ACCELERATORS

PART NUMBER:

TCSK40M-PB

TESLA K40 GPU COMPUTING ACCELERATOR

Solve your most demanding High-Performance Computing (HPC) challenges with NVIDIA Tesla family of GPUs. They’re built on
the NVIDIA Kepler™ compute architecture and powered by NVIDIA CUDA®, the world’s most pervasive parallel computing model.
This makes them ideal for delivering record acceleration and more efcient compute performance for big data applications
in elds, including seismic processing; computational biology and chemistry; weather and climate modeling; image, video and
signal processing; computational nance, computational physics; CAE and CFD; and data analytics.

Tesla K40 GPU Accelerator

Equipped with 12 GB of memory, the Tesla K40 GPU accelerator is ideal for the most demanding HPC and big data problem sets.
It outperforms CPUs by up to 10x 2 and includes a Tesla GPUBoost3 feature that enables power headroom to be converted into
usercontrolled performance boost.

The innovative design of the Kepler compute architecture includes:

>> SMX (streaming multiprocessor)

Delivers up to 3x more performance per watt than the SM in last-generation NVIDIA Fermi GPUs1.

>> Dynamic Parallelism

Enables GPU threads to automatically spawn new threads.
By adapting to the data without going back to the CPU, this greatly simplies parallel programming.

>> Hyper-Q

Allows multiple CPU cores to simultaneously use the CUDA cores on a single Kepler GPU.
This dramatically increases GPU utilization and slashes CPU idle times.

TESLA K40 Module - PRODUCT SPECIFICATIONS

CUDA PARALLEL PROCESSING CORES 2880

FRAME BUFFER MEMORY 12 GB GDDR5

PEAK DOUBLE PRECISION FLOATING POINT PERFORMANCE 1.43 Tops

PEAK SINGLE PRECISION FLOATING POINT PERFORMANCE 4.29 Tops

INTERFACE 384-bit

MEMORY BANDWIDTH 288 GB/s

DISPLAY CONNECTORS None

MAX POWER CONSUMPTION 235 W

PROCESSOR CORE CLOCK 745 MHz

POWER CONNECTORS

GRAPHICS BUS PCI Express 3.0 x16

FORM FACTOR 110 mm (H) × 265 mm (L) - Dual Slot, Full-Height

THERMAL SOLUTION Passive

More information: www.pny.eu/tesla

1 × 6-pin PCI Express power connectors
1 × 8-pin PCI Express power connectors

[e-mail] | T: [phone]

PNY Technologies Europe

NVIDIA® TESLA® K40 Workstation Card

Tesla GPU Computing Accelerator Common Features

ECC MEMORY ERROR PROTECTION

datacenters and supercomputing centers. Both external and internal
memories are ECC protected in Tesla K40.

Integrates the GPU subsystem with the host system’s monitoring and

Meets a critical requirement for computing accuracy and reliability in

SYSTEM MONITORING FEATURES

management capabilities such as IPMI or OEM-proprietary tools. IT
staff can thus manage the GPU processors in the computing system
using widely used cluster/grid management solutions.

Accelerates algorithms such as physics solvers, ray-tracing, and

L1 AND L2 CACHES

sparse matrix multiplication where data addresses are not known
beforehand

ASYNCHRONOUS TRANSFER WITH DUAL DMA
ENGINES

FLEXIBLE PROGRAMMING ENVIRONMENT WITH
BROAD SUPPORT OF PROGRAMMING LANGUAGES
AND APIS

TESLA GPUBoost

Turbocharges system performance by transferring data over the PCIe
bus while the computing cores are crunching other data.

Choose OpenACC, CUDA toolkits for C, C++, or Fortran to express
application parallelism and take advantage of the innovative Kepler
architecture.

End-user can convert power headroom to higher clocks and achieve
even greater acceleration for various HPC workloads on Tesla K40.

Software and Drivers

Software applications page www.nvidia.com/teslaapps

Tesla GPU computing accelerators are supported for both Linux and Windows.

Drivers www.pny.eu/drivers

Learn more about Tesla data center management tools at www.nvidia.com/softwarefortesla

Technical Specications

TESLA K40 TESLA K20X TESLA K20 TESLA K10

Peak double-precision oating point
performance (board)

Peak single-precision oating point
performance (board)

Number of GPUs 1 x GK110B 1 x GK110 2 x GK104s

Number of CUDA cores 2880 2688 2496 2 x 1536

Memory size per board (GDDR5) 12GB 6GB 5GB 8GB

Memory bandwidth for board (ECC off)

2

Architecture features SMX, Dynamic Parallelism, Hyper-Q SMX

System

1.43 Tops 1.31 Tops 1.17 Tops 0.19 Tops

4.29 Tops 3.95 Tops 3.52 Tops 4.58 Tops

288 GB/s 250 GB/s 208 GB/s 320 GB/s

Servers &
workstations

Servers

Servers &
workstations

Servers

4

1. Based on DGEMM performance: Tesla M2090 = 410 gigaops, Tesla K40 > 1000 gigaops

2. Based on SPECFEM3D performance comparison between single E5-2687W @ 3.20GHz vs single Tesla K40

3. For details on GPUBoost refer to the K40 Board spec

4. Tesla K10 specications are shown as aggregate of two GPUs.

More information: www.pny.eu/tesla

PNY Technologies Europe

[e-mail] | T: [phone]