May 2016 Summaries

Denver-based Boom Technology is utilizing Rescale's cloud-based simulation and optimization platform to revitalize supersonic passenger travel with a significantly more efficient and affordable aircraft than the Concorde. Modern computational fluid dynamics (CFD) simulations, enabled by Rescale's platform, allow Boom to validate their aircraft's design and fuel efficiency targets in a matter of hours, a stark contrast to the months required by traditional wind tunnel testing. This approach not only saves Boom substantial costs on server hardware but also provides the flexibility to handle unpredictable demands for computational resources, crucial for meeting engineering schedules and time-to-market objectives. By leveraging Rescale's elastic cloud HPC resources, Boom can accelerate development and minimize capital expenditures, aligning with their goal of making supersonic travel routine and economically viable.

Future Facilities has partnered with Rescale to offer cloud-based high-performance computing (HPC) capabilities to users of its 6SigmaET and 6SigmaDCX computational fluid dynamics (CFD) software, enabling facility managers, architects, and designers to conduct complex data center modeling efficiently. This collaboration allows customers to utilize Rescale’s scalable and secure cloud resources, which can be used when on-premise IT infrastructure is insufficient or for exclusive cloud-based simulations, thus eliminating high initial hardware costs. Jonathan Leppard from Future Facilities highlights the need for sophisticated modeling due to the growing complexity and demand for data centers driven by the proliferation of internet-of-things devices. The partnership aims to balance the speed and accuracy of simulations by leveraging Rescale's computing capabilities, which Joris Poort of Rescale emphasizes will support the development of next-generation devices and data centers. Both companies plan to demonstrate the integration of their products through a joint webinar, showcasing how to efficiently set up and run simulations on Rescale’s platform.

React version 15 has deprecated the use of the valueLink property, which was previously utilized for two-way binding between form element values and component state properties. The recommended approach now involves explicitly specifying the value as a prop and providing an onChange handler. At Rescale, where the use of valueLink was prevalent, this change led to the adoption of a custom implementation using a factory function to generate onChange handlers. This approach supports both simple component states and states referencing immutable objects, eliminating the need for mixins like react-addons-linked-state-mixin, which may face deprecation. The custom solution enhances flexibility by allowing for easy comparison, reversion, or abandonment of changes, while also facilitating deep modifications of immutable objects without relying on deprecated dependencies.

Mark Whitney's article explores two approaches to optimize hyper-parameters in deep neural networks (DNNs) using Rescale's platform. The first method involves a randomized search through the Design-of-Experiments (DOE) framework and demonstrates the process using a convolutional network to classify MNIST digits, varying parameters like the number of filters and convolutional kernel size. The second approach leverages the Sequential Model-based Algorithm Configuration (SMAC) optimizer to systematically explore hyper-parameter configurations, considering additional parameters such as dropout fractions and pooling layer size. Both methods aim to enhance model performance, with the randomized search yielding a slight accuracy improvement and the SMAC optimizer offering a structured way to probe parameter space. The article provides detailed steps for setting up and executing these optimization jobs on Rescale, illustrating the potential for improved neural network performance through strategic hyper-parameter tuning.

TensorFlow, an open-source machine learning library released by Google, has gained significant attention for its accessibility, even to those with limited machine learning experience, and is now available on the Rescale platform, allowing users to create and train models using just a web browser. The blog post guides users through running TensorFlow's introductory tutorial on MNIST using Rescale's GPU hardware, detailing the steps from account creation to job submission, with an example script achieving 91.45% accuracy. It also mentions a more advanced model using a multilayer convolutional network, achieving 99.32% accuracy, and explores performance benchmarks comparing single and multiple GPU setups, highlighting that using four GPUs offers only 2.37 times the performance of a single GPU. The post notes TensorFlow's latest distributed version, which enables workload distribution across multiple GPUs and machines, and hints at future developments to simplify launching multi-node-multi-GPU clusters on Rescale.