Recently we’ve been evaluating various deep learning frameworks including Torch, Caffe, and Tensorflow to understand how they perform and scale be it on CPUs or GPUs. While evaluating performance across a vast array of hardware, ranging from singe nodes with low end GPUs to multi-node Titan X and K80 clusters featuring InfiniBand, we found that more often than note we had to build VM or bare-metal images to get up and running. As you can imagine, building the VMs, while not exactly a complex task, can be vastly time consuming particularly when one has to sort out incompatibilities between Nvidia driver versions, Cuda versions, cuDNN versions, and application versions. Long story short, we realized that we would be better off formalizing this process into tested, reliable, and optimized images which can be easily built and updated whenever there are drive updates, application updates, Cuda updates and so on.
As we enter the Thanksgiving holiday, I wanted to take a moment and write this note to thank our extraordinary Bitfusion team, friends, family, mentors, advisors, investors and our customers, who have been along side with us on our amazing journey in the past year.
As part of the Super Computing conference in Austin, TX last week we added various compute heavy workloads to the Profiler Benchmark Repository. We added application benchmarks in the following categories: Machine Learning, Molecular Dynamics, Fluid Dynamics, Quantum Chemistry. Below is a list of the applications and their respective categories:
Bitfusion extends a warm welcome to SC15, which is returning to Austin after six years. We are proud to participate at SuperComputing 15 and will be showing live demos of our acceleration technology speeding up applications in Scientific Computing, Image Processing, Machine Learning by an order or magnitude, at our booth (#2708).
A little more than a week ago we launched Bitfusion FastR out of Bitfusion Labs, a fully managed R in the cloud service which features various speed enhancement over typical R distributions as well as an online IDE for working with R directly in your browser. Working on this new service while receiving user feedback we learned about quite a few things which we could also add to our AMI sto make them more appealing for our existing customers and new user. Today we are releasing version 0.07 of our Accelerated Scientific Computing AMI on AWS along with the following enhancements:
Since releasing the Bitfusion Profiler a couple weeks ago we've been getting a steady stream of customer feedback from our users. We've been working on the user interface to make it more intuitive and on several back-end tweaks to improve performance. One of the most requested features was to implement multi-cloud support which gives users the ability to benchmark and profile their applications across instances from various cloud providers. Today, we are launching multi-cloud support and have incorporated several instances from Rackspace into Bitfusion Profiler, along with over 20 instance types from AWS. The updated report screen looks as follows:
Today we are releasing Bitfusion FastR as the next product out of Bitfusion Labs in a limited beta. Bitfusion FastR, an online SaaS solution with our proprietary acceleration technology baked-in, is for anyone who needs better performance for their R workloads. FastR offers up to 2x faster run-time performance over typical R distributions across a whole range of programs and benchmarks. We are offering the FastR service in various flavors, starting with a Free plan if you just wanting get your feet wet with our offering or if you are new to R programming, all the way to enterprise grade plans which offer powerful machines that feature dozens of cores and vast amounts of memory for big data workloads.
Last week we released the Bitfusion Profiler as a first product out of Bitfusion Labs. Profiler enables one to quickly evaluate application performance across a variety of hardware and software configurations. It automatically detects limitations and whether a particular application might benefit from larger memory footprints, multiple sockets, more cores, larger disk drives or even a different cloud provider. Then Profiler suggests optimal configurations, helping users determine which instance types are fastest and which offer the best value. We built Profiler initially to meet our internal needs, and now because of customer demand, have put it out for everyone to use. Below is a short overview of how one can get started with the Profiler including an overview of the functionality and a quick tutorial of how one can get started quickly by modifying a sample application workload.
Today, I'm proud to announce that we are launching Bitfusion Labs, a “collaborative proving ground for our research and development efforts.” Labs is the new vehicle by which we will bring platform solutions to market.
Many key areas—pharmaceuticals, data analytics, deep learning, financial services—require complex computations to improve turn-around time to speed-up time-to-market and boost company profits. The volume of data and demand for compute continues to mount and the only only recourse has been to keep scaling: adding CPU-based server nodes. Scale-out definitely helps, but it also means ever lower efficiency as more nodes are added, increased complexity from managing many nodes, increased expenses, and ultimately higher response times overall.