News

Software Solutions Aim to Accelerate AI, Quantum, and GPU Processing

August 26, 2022 by Ikimi .O

With the goal of accelerating various types of advanced computing, including AI, GPU, and quantum processing, a variety of software solutions have recently been unveiled.

A crop of novel software innovations have emerged from leading processor vendors and academia. These solutions promise significant improvements to various applications, including artificial intelligence (AI), high-performance computing (HCP), health care, manufacturing, and more.

These innovations offer several benefits to system developers and EEs, especially in bridging the gap within software-hardware integration and accelerating AI, quantum processing, and GPU systems.

In this article, we’ll provide an overview of these innovations, exploring their impact on modern-day computing.

 

Programming Hardware Accelerators with Exocompilation 

In a recently published paper, MIT researcher Yuka Ikarashi, and his team of researchers from MIT and UC Berkeley, developed and demonstrated the capability of a new programming language that enables the high-performance coding of hardware accelerators.

 

Overview of the Exo system.

Overview of the Exo system. Image used courtesy of Ikarashi and co-authors

 

Called Exo, this new programming language dramatically minimizes the workload of engineers and developers tasked with configuring out-of-box hardware accelerators with software to optimize compatibility with entire application systems. Moreover, Exo allows low-level performance engineers to achieve complex program transformation from simple programs at unprecedented speeds using special accelerator chips.

The language can leverage these accelerators in a process called “Exocompilation” to convert a simple matrix multiplication into a complex program and run it at high speeds.

 

Exocompilation Prioritizes Performance

Unlike existing studies that aim at automating the optimization process for relevant hardware, which, in turn, leads to little to zero process improvements for performance engineers, exocompilation ensures performance engineers prioritize system performance improvements over complex code debugging and optimizations.

With Exo, engineers can isolate and externalize hardware-specific backends (or optimizations) from the compiler and achieve system efficiencies up to 90%. The externalization better separates the open-source project Exo programming language and the proprietary hardware-specific code.

Although Exo leverages specific hardware accelerator chips, the team identified the possibility of expanding its semantics to support parallel programming models and to apply it to even more accelerators, including graphics processing units (GPUs).

 

Intel Rolls Out Open Source AI Reference Kits 

In its bid to contribute to the growing demand for AI to power several sectors, such as agriculture, automotive, electronics, manufacturing, and others, Intel recently released its first set of open source AI reference kits.

In addition to prioritizing accuracy, performance, and low costs, Intel says it developed these kits to enable data scientists and developers to achieve faster and more simplified deployments of AI across retail, healthcare, and other productivity-intensive industries.

Intel’s new open source AI kits include:

  • Utility asset health
  • Visual quality control
  • Customer chatbot
  • Intelligent document indexing

Next, we’ll discuss each kit in detail.

 

Utility Asset Health Reference Kit

The Utility Asset Health reference kit is a predictive analytics model that ensures higher service reliability in utilities. This kit meets the need for growth in power distribution assets resulting from the underlying global hike in energy consumptio, says Intel.

Utility Asset Health models the health of utility poles through the Intel one API Data Analytics Library. Asset age, mechanical properties, manufacturer, prior repair, geospatial data, inspections, etc., are some of the data the model uses for its analysis.

 

Utility Asset Health reference kit.

Utility Asset Health reference kit. Image used courtesy of Intel

 

Visual Quality Control Reference Kit

Acknowledging the critical nature of quality control (QC) in manufacturing operations and the heavy graphics requirements of computer vision QC techniques, Intel released its Visual Quality Control reference kit.

 

Intel Visual QC reference kit.

Intel Visual QC reference kit. Image used courtesy of Intel

 

The AI visual QC model integrates Intel Optimization for PyTorch and Intel Distribution of OpenVINO toolkit with oneAPI to achieve faster training and inferencing in QC applications. According to Intel, the Visual Quality Control reference kit can deliver up to 95% accuracy in hyperparameter tuning and optimization-based pharmaceutical pill defect detections.

 

Customer Chabot Reference Kit

The Customer Chabot reference kit ensures the deployment of AI to support massive and highly complex conversational chatbot interactions, says Intel. This kit utilizes Intel Extension for PyTorch and Intel Distribution of OpenVINO toolkit to optimize the AI model for better performance. The reference kit also minimizes model development code modifications for training and inferencing.

 

Customer Chatbot reference kit.

Customer Chatbot reference kit. Image used courtesy of Intel

 

Intelligent Document Indexing Reference Kit

With the Intelligent Document Indexing reference kit, enterprises can rapidly deploy AI to process and analyze faster at reduced labor costs. The kit leverages the integration of a support vector classification (SVC) model with Intel Distribution of Modin and Intel Extension for Scikit-learn powered by oneAPI to improve training, inferencing, and data processing times in indexing and sorting documents at up to 95% accuracy, says Intel.

 

Digitized document routing with Intelligent Document Indexing.

Digitized document routing with Intelligent Document Indexing. Image used courtesy of Intel

 

Nvidia’s QODA Blends Quantum and Classical Computing

Nvidia is fostering the widespread adoption and commercialization of quantum computing. Along those lines, in a July announcement, the company introduced its Quantum Optimized Device Architecture (or QODA), which harnesses a coherent hybrid quantum-classical programming model that promises scientific productivity improvements and expands the quantum research scale.

Moreover, the architecture can enable significant advancements in conventional applications, increasing the potential for unprecedented scientific breakthroughs in the near term. Although Nvidia says that top quantum organizations are already using Nvidia GPUs and highly specialized Nvidia software (Nvidia cuQuantum) to develop standalone quantum circuits, QODA can allow developers to build complete quantum applications on GPU-accelerated supercomputers.

 

QODA offers a unified programming model designed for quantum processors in a hybrid setting, working alongside CPUs and GPUs.

QODA offers a unified programming model designed for quantum processors in a hybrid setting, working alongside CPUs and GPUs. Image used courtesy of Nvidia

 

According to Nvidia, QODA offers flexibility and scalability, openness, high performance, easy integration, and high productivity benefits to several applications. The architecture promises high flexibility and scalability by supporting hybrid deployments via emulation on a single GPU up to Nvidia DGX SuperPOD plus multiple quantum processing unit (QPU) partner backends.

Because it’s open source, QODA can connect to various QPU backend types and allow access to all users. The technology provides high performance with speedups of up to 287× in end-to-end Variational Quantum Eigensolver (VQE). OODA also offers interoperability with modern GPU-accelerated applications, and improved productivity in quantum algorithm research due to a streamlined hybrid quantum-classical development. All that is expected to help HPC developers achieve accelerated applications in a consolidated environment.

 

New Tools for a New Era

Gone now are the days when electronic systems designers only had traditional CPU-based computing to consider. While AI and GPU computing aren’t new, in the grand scale of things, they are relatively young in terms of the kinds of software solutions available for accelerating systems.

Quantum computing meanwhile is definitely in its nascent stages. To serve the needs of all these types of computing, the solutions discussed here are perhaps a step toward making all  these technologies easier to implement  for engineers.