ATS-GPU-BASE
5Pages

Real Time Signal Processing Software • Transfer A/D data to GPU at high speed • Up to 6.9 GB/s transfer rate for PCIe Gen 3 digitizer boards ATS-GPU Based User Application Control Data Flow • Supports CUDA -enabled GPUs with compute capability 3.0 and higher‡ Custom Kernels • Designed to work with AlazarTech PCI Express waveform digitizers • Availability of an optional OCT Signal Processing Library that includes: CUDA Runtime AlazarTech Driver AlazarTech Digitizer Card ◊ Very high-speed floating-point FFT ◊ Dispersion compensation and windowing functions NVIDIA Driver • Compatible with Windows® & Linux® Product Operating System Kernel Programming Language CUDA compute capability 3.0+‡ second, demonstrating the power and efficiency of the ATSGPU‑BASE platform. Kernel code running on the GPU can apply a windowing function, do a floating point FFT, calculate the amplitude, and convert the result to a log scale. Please refer to the ATS‑GPU-OCT datasheet for more information. Interfacing waveform digitizers to GPUs involves creating a software mechanism to move data from one to the other and back to user buffers. The standard techniques used most often can get the job done, but feature very low data throughput due to software overheads. ATS-GPU-BASE Limitations: ATS-GPU-BASE is a software library developed by AlazarTech that transfers data acquired by its family of PCI Express waveform digitizers to a CUDA-enabled Graphical Processing unit (GPU) at sustained transfer rates as high as 6.9 GB/s. AlazarTech designed ATS-GPU to eliminate this software bottleneck so that data can be moved from AlazarTech digitizers to GPUs and from GPUs to user buffers at full PCIe bus speeds. Once the data is available in GPU memory, many types of digital signal processing (DSP) can be done on this data at near-hardware speeds. Data transferred by ATS-GPU-BASE is presented in GPU memory as a buffer queue. Expert-level GPU programmers can create very high-performance custom kernels to manipulate this data using an easy-to-use application programming interface (API). ATS-GPU-BASE includes an example program that demonstrates how to use the ATS-GPU library to transfer data from a waveform digitizer to a GPU. The example also shows how to do simple data processing on the GPU using CUDA kernels, and how to transfer the processed data back to host memory (RAM). Users can use this example program as a starting point to create their own custom kernels to do GPU-based DSP. An example of a high-performance kernel is the optional ATS‑GPU-OCT library (sold separately), which acquires data using ATS‑GPU‑BASE and then implements full OCT signal processing algorithm using CUDA kernels. ATSGPU-OCT was benchmarked at up to 950,000 4K FFTs per Page 1 Technical Support: Users should note that technical support for ATS-GPU-BASE is limited to the published user manual; no other technical support will be provided. Programming language: Note that ATS-GPU-BASE development can only be done using C/C++; it is not possible to program in Python, MATLAB, or LabVIEW. Required programming knowledge: Users must have expert programming knowledge of CUDA development in order to customize ATS-GPU kernels. Use-case approval requirement: In order to avoid possible disappointment in product functionality and to ensure that customers are selecting the correct solution, AlazarTech requires that customers complete a questionnaire prior to purchasing ATS-GPU-BASE as a standalone library. GPU-Based Signal Processing Graphical Processing Units (GPUs) were originally designed for rendering high-quality video for gaming applications, which required being able to perform massive amount of real-time calculations. The highly parallel architecture of modern GPUs also makes them an ideal platform for digital signal processing (DSP) and high performance computing (HPC) systems. In the past, complex real-time signal processing, such as FFT, correlation, FIR filtering etc., could only be achieved using dedicated DSP processors or by implementing the algorithms inside an FPGA or an ASIC. All these methods are non-trivial, expensive, time consuming and require highly specialized engineering skills.

Real Time Signal Processing Software Using GPUs, users can implement any algorithm that can be parallelized in a GPU using well known software techniques and gain a better than 10-fold improvement over CPU based signal processing. The reason why GPUs perform so well for DSP applications is that they contain hundreds of processing cores (kernels) running in parallel, while sharing a very high speed graphical memory bank. ATS-GPU-BASE uses multiple CUDA streams to move data between the digitizer and GPU. This means there is a latency between data being acquired by the digitizer board and GPU...

Real Time Signal Processing Software as the host CPU is more than capable of handling data rates generated by PCI bus boards. ATS-GPU cannot directly be interfaced with non-AlazarTech waveform digitizers. The data transfer rate to GPU is dependent on the generation of PCI Express digitizer board used or Thunderbolt 3 system configuration: PCIe Link Speed Transfer Rate Thunderbolt 3: ATST371, ATST146, ATST352 *dependent on system configuration project and a CMake build file. We use more recent C++ features, and Visual Studio 2015 and later is required. On Linux, a C++11 compiler is required...

Real Time Signal Processing Software ORDERING INFORMATION ATS-GPU-BASE: GPU Streaming Library ATS-GPU-BASE-1YR: 1 year extended maintenance for ATS-GPU-BASE ATS-GPU-OCT: Signal Processing Library ATS-GPU-OCT-1YR: 1 year extended maintenance for ATS-GPU-OCT License + 1 Year Subscription License + 1 Year Subscription (requires ATSGPU-001; also requires ATS-SDK for use with Python, MATLAB, & LabVIEW) ATS-GPU-NUFFT: ATS-GPU-OCT Extension ATSGPU-201 for fixed-frequency sampled data License + 1 Year Subscription (requires ATSGPU-001 and ATSGPU-101) ATS-GPU-NUFFT-1YR: 1 year extended maintenance...