12 GTC 2015 Sessions Not to Miss

With one week to go until we all descend on GTC 2015, I’ve scoured through the list of Accelerated Computing sessions and put together 12 diverse “not to miss” talks you should add to your planner. This year, the conference is highlighting the revolution in Deep Learning that will affect every aspect of computing. GTC 2015 includes over 40 session categories, including deep learning and machine learning, scientific visualization, cloud computing, and HPC.

This is the place where scientists, programmers, researchers, and a
myriad of creative professionals convene to tap into the power of a GPU
for more than gaming. –Forbes

Tuesday, March 17

An Introduction to CUDA Programming (S5661)


This is the introductory tutorial intended for those new to CUDA and you will leave with the essential knowledge to start programming in CUDA – no experience is needed! For those that have prior CUDA experience, this is a great session to brush up on key concepts required for subsequent tutorials on CUDA optimization. The other tutorials in this session are: An Introduction to the GPU Memory ModelAsynchronous Operations and Dynamic Parallelism in CUDA and Essential CUDA Optimization Techniques.

GTC attendees learn from the brightest minds in accelerated computing with hundreds of talks and hands-on tutorials.
GTC attendees learn from the brightest minds in accelerated computing with hundreds of talks and hands-on tutorials.

SMTool: A GPU based Satellite Image Analysis Tool (S5201)


Dilip Patlolla, R&D Engineer in the Geographic Information Science and Technology (GIST) Group at the Oak Ridge National Laboratory, will demonstrate their advanced satellite image analytic tool referred as SMTool built on the CUDA platform to process city-scale sub-meter resolution satellite imagery to detect and discriminate man-made structures. Continue reading


Understanding Natural Language with Deep Neural Networks Using Torch

This post was co-written by Soumith Chintala and Wojciech Zaremba of Facebook AI Research.

Language is the medium of human communication. Giving machines the ability to learn and understand language enables products and possibilities that are not imaginable today.

One can understand language at varying granularities. When you learn a new language, you start with words: understanding their meaning, identifying similar and dissimilar words, and developing a sense of contextual appropriateness of a word. You start with a small dictionary of words, building up your dictionary over time, mentally mapping each newly learned word close to similar words in your dictionary. Once you get familiar with your dictionary of words, you put them together into small sentences, learning grammar and structure. You eventually combine sentences in a sensible way, to write paragraphs and pages. Once you get to this stage, you are comfortable with expressing complicated thoughts in language, letting others understand your thoughts and expression.

As an example, language understanding gives one the ability to understand that the sentences “I’m on my way home.” and “I’m driving back home.” both convey that the speaker is going home.

Word Maps and Language Models

For a machine to understand language, it first has to develop a mental map of words, their meanings and interactions with other words. It needs to build a dictionary of words, and understand where they stand semantically and contextually, compared to other words in their dictionary. To achieve this, each word is mapped to a set of numbers in a high-dimensional space, which are called “word embeddings”. Similar words are close to each other in this number space, and dissimilar words are far apart. Some word embeddings encode mathematical properties such as addition and subtraction (For some examples, see Table 1).

Word embeddings can either be learned in a general-purpose fashion before-hand by reading large amounts of text (like Wikipedia), or specially learned for a particular task (like sentiment analysis). We go into a little more detail on learning word embeddings in a later section.

Table 1: Mikolov et. al. [3] showcase simple additive properties of their word embeddings.
Expression Nearest Token
Paris – France + Italy Rome
bigger – big + cold colder
sushi – Japan + Germany bratwurst
Cu – copper + gold Au
Windows – Microsoft + Google Android
Montral Canadiens – Montreal + Toronto Toronto Maple Leafs

After the machine has learned word embeddings, the next problem to tackle is the ability to string words together appropriately in small, grammatically correct sentences which make sense. This is called language modeling. Language modeling is one part of quantifying how well the machine understands language. Continue reading


Deep Speech: Accurate Speech Recognition with GPU-Accelerated Deep Learning

Speech recognition is an established technology, but it tends to fail when we need it the most, such as in noisy or crowded environments, or when the speaker is far away from the microphone. At Baidu we are working to enable truly ubiquitous, natural speech interfaces. In order to achieve this, we must improve the accuracy of speech recognition, especially in these challenging environments. We set out to make progress towards this goal by applying Deep Learning in a new way to speech recognition.

Figure 1: The structure of our deep neural network, showing the layers (top to bottom) and how we parallelize training across GPUs (left to right). The fourth layer is a bidirectional recurrent layer. Blue and red arrows indicate the forward and backward direction and the communication required between GPUs in these layers.
Figure 1: The structure of our deep neural network, showing the layers (top to bottom) and how we parallelize training across GPUs (left to right). The fourth layer is a bidirectional recurrent layer. Blue and red arrows indicate the forward and backward direction and the communication required between GPUs in these layers.

Deep Learning has transformed many important tasks; it has been successful because it scales well: it can absorb large amounts of data to create highly accurate models. Indeed, most industrial speech recognition systems rely on Deep Neural Networks as a component, usually combined with other algorithms. Many researchers have long believed that Deep Neural Networks (DNNs) could provide even better accuracy for speech recognition if they were used for the entire system, rather than just as the acoustic modeling component. However, it has proven difficult to find an end-to-end speech recognition system based on Deep Learning that improves on the state of the art.

Model and Data Co-design

One of the reasons this has been difficult is that training these networks on large datasets is computationally very intensive. The process of training DNNs is iterative: we instantiate ideas about models in computer code that trains a model, then we train the model on a training set and test it, which gives us new ideas about how to improve the model or training set. The latency of this loop is the rate limiting step that gates progress. Our models are relatively large, containing billions of connections, and we train them on thousands of hours of data, which means that training our models takes a lot of computation. Continue reading

GTC attendees learn from the brightest minds in accelerated computing with hundreds of talks and hands-on tutorials.

Learn GPU Computing with Hands-On Labs at GTC 2015

Every year NVIDIA’s GPU Technology Conference (GTC) gets bigger and better. One of the aims of GTC is to give developers, scientists, and practitioners opportunities to learn with hands-on labs how to use accelerated computing in their work. This year we are nearly doubling the amount of hands-on training provided from last year, with almost 2,400 lab hours available to GTC attendees!

We have two types of training this year at GTC: instructor-led labs and self-paced labs. And to help you keep up with one of the hottest trends in computing, this year we’re featuring a Deep Learning training track. Keep reading for details. If you haven’t registered for GTC yet this year, keep reading for a discount code.

Deep Learning Track

There is an explosion of Deep Learning topics at GTC, and it’s not limited to the keynotes, talks and tutorial sessions. We’ll feature at least six hands-on labs related to accelerating facets of Deep Learning on GPUs. From an introduction to Deep Learning on GPUs to cutting-edge techniques and tools, there will be something for everyone. Be sure to get to these labs early to get yourself a seat! Here are a few of the labs available in this track:

  • Introduction to Machine Learning with GPUs: Handwritten digit classification (S5674)
  • DIY Deep Learning for Vision with Caffe (S5647)
  • Applied Deep Learning for Vision, Natural Language and Audio with Torch7 (S5574)
  • Deep Learning with the Theano Python Library (S5732)
  • Deep Belief Networks Using ArrayFire (S5722)
  • Accelerate a Machine Learning C++ example with Thrust (S5822)

Instructor-led Labs

IMAG0568Just like GTC last year, there will be twenty hands-on instructor-led labs. These are 80-minute labs led by an expert on the topic. Continue reading

BIDMach: Machine Learning at the Limit with GPUs

Deep learning has made enormous leaps forward thanks to GPU hardware. But much Big Data analysis is still done with classical methods on sparse data. Tasks like click prediction, personalization, recommendation, search ranking, etc. still account for most of the revenue from commercial data analysis. The role of GPUs in that realm has been less clear. In the BIDMach project (part of the BID Data Project at UC Berkeley), we have been exploring general machine learning with GPUs. The results are remarkable: not only do we see order-of-magnitude speedups for most problems, but our system also outperforms today’s cluster computing systems running up to several hundred nodes on typical workloads. As well as the incentives to adopt GPU technology for deep learning tasks, there is now a strong incentive for organizations to migrate to GPUs for the remainder of their analytics workload.

Roofline Design

To build the fastest system, we borrowed the approach of roofline design from computer architecture. Roofline design involves designing to fundamental limits (e.g. ALU throughput, memory speed, network speed, I/O speed etc). A rooflined system is fast, and no other system can be much faster, since both have to respect the same hardware limits. A roofline diagram for matrix multiply is shown below:

Figure 1: CPU and GPU roofline limits
Figure 1: CPU and GPU roofline limits

The y-axis shows the potential throughput in arithmetic operations/second. The x-axis is “operational intensity” which is the number of operations applied to each data value (in units of operations per byte). The intensity is much lower for sparse operations – e.g. sparse matrix multiply typically involves only a multiply and add for each input datum, while dense matrix multiply uses each datum many times. The horizontal lines reflect the maximum ALU throughput for each type of processor (the graph is drawn for Intel i7 and NVIDIA GeForce GTX 680 processors). GPUs have much higher ALU throughput since the GPU chip area is almost entirely ALU. For dense matrix multiply, GPUs are10x faster thanks to this higher computing area.

The diagonal lines reflect memory bandwidth. Since bandwidth is flow in bytes/second it defines a linear relationship between the x-axis (flops/byte) and the y-axis (in flops/sec). A less well-known feature of GPUs is their higher main-memory bandwidth. This leads to a (potential) 10x gap in sparse matrix operations, which are the most important for many machine learning tasks. CuSparse achieves this ceiling for typical scientific data, but we found it was less well fit to very sparse data (text, web logs etc.). We wrote our own sparse kernels and were able to get them close to the roofline limits over a full range of sparseness. These kernels form the basis for the high throughput in most of BIDMach’s algorithms. Continue reading


Accelerating Bioinformatics with NVBIO

NVBIO is an open-source C++ template library of high performance parallel algorithms and containers designed by NVIDIA to accelerate sequence analysis and bioinformatics applications. NVBIO has a threefold focus:

  1. Performance, providing a suite of state-of-the-art parallel algorithms that offer a significant leap in performance;
  2. Reusability, providing a suite of highly expressive and flexible template algorithms that can be easily configured and adjusted to the many different usage scenarios typical in bioinformatics;
  3. Portability, providing a completely cross-platform suite of tools, that can be easily switched from running on NVIDIA GPUs to multi-core CPUs by changing a single template parameter.

Exponential Parallelism

We built NVBIO because we believe only the exponentially increasing parallelism of many-core GPU architectures can provide the immense computational capability required by the exponentially increasing sequencing throughput.

There is a common misconception that GPUs only excel at highly regular, floating point intensive applications, but today’s GPUs are fully programmable parallel processors, offering superior memory bandwidth and latency hiding characteristics, and R&D efforts at NVIDIA and elsewhere have proved that they can be a perfect match even for branchy, integer-heavy bioinformatics applications. The caveat is that legacy applications need to be rethought for fine-grained parallelism.

Many CPU algorithms are designed to run on few cores and scale to a tiny number of threads. When the number of threads is measured in the thousands, rather than dozens—a fact that all applications inevitably must consider—applications must tackle fundamental problems related to load balancing, synchronization, and execution and memory divergence.

NVBIO does just that, providing both low-level primitives that can be used from either CPU/host or GPU/device threads, as well as novel, highly parallel high-level primitives designed to scale from the ground up. Continue reading


GPU Pro Tip: Fast Dynamic Indexing of Private Arrays in CUDA

Sometimes you need to use small per-thread arrays in your GPU kernels. The performance of accessing elements in these arrays can vary depending on a number of factors. In this post I’ll cover several common scenarios ranging from fast static indexing to more complex and challenging use cases.

Static indexing

Before discussing dynamic indexing let’s briefly look at static indexing. For small arrays where all indices are known constants at compile time, as in the following sample code, the compiler places all accessed elements of the array into registers.

__global__ void kernel1(float * buf)
    float a[2];
    float sum = a[0] + a[1];

This way array elements are accessed in the fastest way possible: math instructions use the data directly without loads and stores.

A slightly more complex (and probably more useful) case is an unrolled loop over the indices of the array. In the following code the compiler is also capable of assigning the accessed array elements to registers.

__global__ void kernel2(float * buf)
    float a[5];
    float sum = 0.0f;
    #pragma unroll
    for(int i = 0; i < 5; ++i)
        sum += a[i];

Here we tell the compiler to unroll the loop with the directive #pragma unroll, effectively replacing the loop with all the iterations listed explicitly, as in the following snippet.

sum += a[0];
sum += a[1];
sum += a[2];
sum += a[3];
sum += a[4];

All the indices are now constants, so the compiler puts the whole array into registers. Continue reading


GPU Pro Tip: CUDA 7 Streams Simplify Concurrency

Heterogeneous computing is about efficiently using all processors in the system, including CPUs and GPUs. To do this, applications must execute functions concurrently on multiple processors. CUDA Applications manage concurrency by executing asynchronous commands in streams, sequences of commands that execute in order. Different streams may execute their commands concurrently or out of order with respect to each other. [See the post How to Overlap Data Transfers in CUDA C/C++ for an example]

When you execute asynchronous CUDA commands without specifying a stream, the runtime uses the default stream. Before CUDA 7, the default stream is a special stream which implicitly synchronizes with all other streams on the device.

CUDA 7 introduces a ton of powerful new functionality, including a new option to use an independent default stream for every host thread, which avoids the serialization of the legacy default stream. In this post I’ll show you how this can simplify achieving concurrency between kernels and data copies in CUDA programs.
Continue reading


CUDA 7 Release Candidate Feature Overview: C++11, New Libraries, and More

It’s almost time for the next major release of the CUDA Toolkit, so I’m excited to tell you about the CUDA 7 Release Candidate, now available to all CUDA Registered Developers. The CUDA Toolkit version 7 expands the capabilities and improves the performance of the Tesla Accelerated Computing Platform and of accelerated computing on NVIDIA GPUs.

Recently NVIDIA released the CUDA Toolkit version 5.5 with support for the IBM POWER architecture. Starting with CUDA 7, all future CUDA Toolkit releases will support POWER CPUs.

CUDA 7 is a huge update to the CUDA platform; there are too many new features and improvements to describe in one blog post, so I’ll touch on some of the most significant ones today. Please refer to the CUDA 7 release notes and documentation for more information. We’ll be covering many of these features in greater detail in future Parallel Forall posts, so check back often!

Support for Powerful C++11 Features

C++11 is a major update to the popular C++ language standard. C++11 includes a long list of new features for simpler, more expressive C++ programming with fewer errors and higher performance. I think Bjarne Stroustrup, the creator of C++, put it best:

C++11 feels like a new language: The pieces just fit together better than they used to and I find a higher-level style of programming more natural than before and as efficient as ever.
Continue reading


How We Achieved Record Finance Benchmark Performance on Tesla K80

STAC Research develops financial benchmarks in partnership with leading banks and software or hardware vendors. The STAC-A2 suite of benchmarks aims to represent the standard risk analysis workload that banks and insurance companies use to measure exposure on the financial markets. Earlier this year we published a Parallel Forall post on Monte Carlo simulation for the pricing of American options in STAC-A2.

Record Performance with Tesla K80

Recently, STAC Research published astonishing performance results for the STAC-A2 benchmarks on an NVIDIA Tesla K80. In short,  a single Tesla K80 driven by two CPU cores outperforms all previously audited systems in terms of pure performance and power efficiency.

For more on these results, read “Bank on It: Tesla Platform Shatters Record on Risk-Management Benchmark” on the NVIDIA Blog.

tesla-k80-3qtrWe obtained these new results after several optimizations of our previously audited code. First of all, a large fraction of the computations are now avoided due to a better factorization of the underlying mathematical process. Secondly, we tuned some of the kernel parameters to take advantage of the larger register file of the Tesla K80. Finally, we were able to significantly reduce the latency in one of the main loops of the benchmark. Let’s take a look at these optimizations. Continue reading