GaGe CompuScope GPU CUDA Processing

www.gage-applied.com

CompuScope GPU CUDA Processing

Real-Time High-Speed Digital Signal Processing

FEATURES

 All CompuScope Express Digitizers with eXpert Streaming Supported

 Stream Acquired Signal Data to GPU for In-Line Processing in Real-Time

 Stream GPU Processed Data Results to Storage in Real-Time

 Up to 5 GB/s Data Stream Rates via PCIe Gen3 Interface

 Supports CUDA Compute Capability 3.0+ GPU Cards

 10X ~ 100X Faster GPU Processing Calculation Times vs CPU Processing

 Easier & More Transportable C Language Based Programming for GPU

 Provided C SDK Sample Programs Illustrate:

o Digitizer Streaming to GPU DSP Algorithms

o Examples Programs such as FFT, Signal Averaging, and More

o Easily Insert User Developed Custom Code for GPU Processing

o Exploitation of GPU Vector Data Processing Capability

o Usage of CUDA Library for Fastest Data Transfer Performance

o Comparative Performance of GPU vs Host CPU Processing

o GPU Streaming of Processed Data Results to Storage

 Available Sig-Station Systems for Maximum Performance

 Windows and Linux Operating Systems Supported

GaGe

is a worldwide industry leader

in high speed data acquisition

solutions featuring a portfolio of the

highest performance digitizers, PC

oscilloscope software, powerful SDKs

for custom application development,

and turnkey integrated PC-based

measurement systems.

APPLICATIONS

Wideband Signal Analysis

RADAR Design and Test

Signals Intelligence (SIGINT)

Ultrasonic Non-Destructive Testing

LIDAR Systems

Communications

Optical Coherence Tomography

Spectroscopy

High-Performance Imaging

Time of Flight

Life Sciences

Particle Physics

www.gage-applied.com

CompuScope GPU CUDA Processing Block Diagram

Overview

Performing real-time digital signal processing (DSP) routines

such as Fast Fourier Transform (FFT), Signal Averaging, Finite

Impulse Response (FIR) Filtering, Digital Down Conversion

(DDC) and more have traditionally required the use of

dedicated DSP processors, Field Programmable Gate Arrays

(FPGAs), or Application Specific Integrated Circuits (ASICs).

The size of DSP processors, FPGAs, and ASICs are often

resource limited and require significant development with

specialized engineering skills in particular languages and

platform tools that are often native and proprietary to the

targeted device. These solutions are typically expensive in

terms of the physical devices, required software development

tools, lengthy development cycles and expertise resources.

With the evolution of increasing multi-core CPU resources,

developers began to implement DSP routines utilizing the

host CPU. The primary advantage of the host CPU platform is

that it enables faster development utilizing more familiar

programming environments such as C; as opposed to more

complex VHDL based programming that is typically required

for FPGAs. However a disadvantage is that the host CPU

platform may still not be powerful enough to conduct

complex processing routines, especially for real-time

application requirements.

The use of a GPU provides the best of both capabilities, a fast

familiar programming development environment such as C

and more powerful computational capabilities than the host

CPU platform to satisfy demanding real-time application

requirements. GPUs incorporate many more cores with a

high-speed memory bus that can be effectively utilized in

parallel for efficient manipulation of large blocks of data.

The PCI Express (PCIe) interface bus is the key pathway for

transferring data to and from the GPU at very high-speeds.

Taking advantage of the GaGe eXpert PCIe Data Streaming

firmware capability, GaGe CompuScope PCIe Digitizer models

can simultaneously acquire and stream data to the GPU at

sustained rates up to 5 GB/s via a PCIe Gen3 interface.

The Gage CompuScope C Software Development Kit (SDK)

provides ready-made compiled sample programs illustrating

how to configure and use the eXpert PCIe Data Streaming

feature to stream acquired data to the GPU for processing

AND how to effectively exploit the parallelized vector

processing of the GPU to attain processing rates anywhere

from 10X ~ 100X faster than the host CPU for display analysis.

This enables end users to quickly and easily begin working

with GPU cards, focusing on the development of their custom

in-line processing routines that is unique to their application.

Projects can be developed rapidly and are more transportable

working in a C programming environment with the GPU CUDA

library.

The Gage CompuScope C SDK sample programs also illustrate

streaming GPU processed data results continuously to high-

speed storage for real-time data recordings of the processed

output. Utilizing PCIe based storage systems allow for

continuous signal capture and processing recording systems

at high-speed sustained transfer rates at up to 5 GB/s within

a Windows based environment.

GPU processed data can also be optionally streamed to PCIe

Arbitrary Waveform Generators to effectively playback

modified signals back out to the testing environment, suitable

for applications such as Digital Radio Frequency Memory

(DRFM).

Input Signal

Display Analysis

Output Signal

Stream to GPU

SSD

RAID

Storage

Digitizer A/D

Stream to Storage

Stream to/from Storage

Stream to AWG

AWG D/A

Stream from Storage

GPU CUDA Processing

www.gage-applied.com

GPU CUDA Series

GaGe supports NVIDIA Tesla, Quadro, and GeForce series of

GPU cards with a CUDA Compute Capability of 3.0+. The

following table provides a general overview of certain

capability ranges of each series for recent 6.0 to 7.5 CUDA

rated models. We recommend checking NVIDIA for detailed

specifications of specific NVIDIA models.

Tesla Series Quadro Series GeForce Series

# of GPU

Cores:

2,560 – 5,120 256 – 5,120 768 – 5,120

# of Tensor

Cores:

320 – 640 288 – 576 240 – 544

GPU

Clock Rates:

0.585 – 1.59

GHz

1.05 – 1.815

GHz

1.2 – 1.785

GHz

Floating Point

64-bit (FP64)

Double-Precision:

0.2 – 7

TFLOPS

0.02 – 7.4

TFLOPS

0.07 – 7.4

TFLOPS

Floating Point

32-bit (FP32)

Single-Precision:

5.5 – 14

TFLOPS

0.6 – 16.3

TFLOPS

2.3 – 16.3

TFLOPS

Floating Point

16-bit (FP16)

Half-Precision:

0.1 – 28

TFLOPS

0.01 – 32.6

TFLOPS

0.07 – 7.45

TFLOPS

Memory

Bandwidth:

192 – 900

GB/s

32 – 870

GB/s

84 – 672

GB/s

GPUs are particularly efficient in performing floating point

calculations by conducting highly parallel computations

across multiple cores simultaneously.

Applications requiring higher accuracy mathematical

calculations can utilize double-precision 64-bit floating point

for data values twice in size than single-precision 32-bit

floating point. Lower half-precision 16-bit floating point can

be utilized for applications that require even less accuracy

such as neural network training or inference. More recently,

Tensor cores have been introduced into GPUs for accelerating

large matrix based operations utilized in deep learning or

artificial intelligence (AI) operations.

CUDA is NVIDIA’s parallel computing platform and

programming model that allows for the GPU to be utilized for

general purpose computing using familiar software

development environments such as C or C++. The CUDA

Toolkit software is provided directly by NVIDIA at no charge

with detailed documentation and various coding samples to

help with application development.

Sig-Station System Options

Optional Sig-Stations are available for providing complete

turn-key systems for GPU capability working with GaGe high-

speed instruments. Sig-Stations are high-performance PC

workstations that are designed specifically for integrating

advanced instruments and maximizing their operational

performance.

For real-time operations, it is critical that the underlying host

platforms are fully capable of sustaining high-speed PCIe data

streaming rates to and from the GPU. Traditional lower cost

desktop based platforms often restrict performance

capabilities by placing multiple PCIe slots behind shared PCIe

switches. Sig-Station systems utilize dedicated bandwidth

PCIe slot architecture for maximum sustained PCIe streaming

rates for multiple cards simultaneously.

Sig-Stations come with all GaGe cards, features, and software

fully tested and installed so that the user can be up and

running with their system solution right out of the box; thus

saving time and minimizing risks of self-integrated systems.

Custom system configurations can be defined to meet specific

customer application requirements.

These systems incorporate the latest in PC-based technology

and utilize workstation class motherboards with multiple

dedicated bandwidth PCIe slots, high multi-core count Xeon

CPUs, and large system memory capacity. Integrated high-

speed data storage systems for real-time signal recording

applications requiring a guaranteed continuous sustained

data streaming rate with no missing data can be included.

www.gage-applied.com

ORDERING INFORMATION

CompuScope GPU CUDA Processing

Order

Part Number

The CompuScope GPU CUDA Processing package includes the following

bundled items:

1.) eXpert PCIe Data Streaming Firmware (STR-181-000)

2.) CompuScope C/C# Software Development Kit (200-200-101)

NOTE: GPU card is not included.

The eXpert PCIe Data Streaming Firmware is compatible for use with the

following GaGe Digitizer Model Series sold separately:

Please refer to the separate GaGe product datasheets for these digitizer

models for their full specification details and ordering information.

The CompuScope C/C# SDK provides ready-made compiled sample

programs illustrating:

 Use of eXpert PCIe Data Streaming feature to stream acquired data

to a GPU for processing.

 How to effectively exploit the parallelized vector processing of the

GPU for signal processing routines with various examples.

 Optionally stream GPU processed data results continuously to high-

speed storage for real-time data recordings of the processed output.

To maximize performance, customers should take care to select suitable

host systems with dedicated bandwidth PCIe slots with sufficient power

and cooling. High-performance GPU cards typically require dual slot space

with additional power and good cooling air flow.

Optional Sig-Stations are available for providing complete turn-key

systems for GPU capability working with GaGe high-speed instruments.

BDL-GPU-000



Cobra Express

 CobraMax Express

 EON Express

 Octave Express

 Octopus Express

 Oscar Express



Razor Express

 RazorPlus Express

 RazorMax Express

900 N. State St.

Lockport, IL 60441-2200

Toll-Free (USA and Canada):

Phone: 1-800-567-4243

Fax: 1-800-780-8411

Direct:

Phone: 1-514-633-7447

Fax: 1-514-633-0770

Email:

[email protected]

To find your local sales representative or

distributor or to learn more about GaGe

products visit:

www.gage-applied.com

WARRANTY

Standard two years parts and labor.

Unless otherwise specified, all dynamic

performance specs have been qualified on

engineering boards. All specifications are

subject to change without notice.

Data Sheet Revision 0 – 03/06/2019

GaGe is a product brand of DynamicSignals

LLC, an ISO 9001:2015 Certified Company