Improving FinanceBench for GPUs Part II – low hanging fruit

We found a finance benchmark for GPUs and wanted to show we could speed its algorithms up. Like a lot!

Following the initial work done in porting the CUDA code to HIP (follow article link here), significant progress was made in tackling the low hanging fruits in the kernels and tackling any potential structural problems outside of the kernel.

Additionally, since the last article, we’ve been in touch with the authors of the original repository. They’ve even invited us to update their repository too. For now it will be on our repository only. We also learnt that the group’s lead, professor John Cavazos, passed away 2 years ago. We hope he would have liked that his work has been revived.

Link to the paper is here:

Scott Grauer-Gray, William Killian, Robert Searles, and John Cavazos. 2013. Accelerating financial applications on the GPU. In Proceedings of the 6th Workshop on General Purpose Processor Using Graphics Processing Units (GPGPU-6). Association for Computing Machinery, New York, NY, USA, 127–136. DOI:

Improving the basics

We could have chosen to rewrite the algorithms from scratch, but first we need to understand the algorithms better. Also, with the existing GPU-code we can quickly assess what are the problems of the algorithm, and see if we can get to high performance without too much effort. In this blog we show these steps.

Continue reading “Improving FinanceBench for GPUs Part II – low hanging fruit”

Question: do we work with CUDA?

Answer: Yes, actually a lot!

The company was built on OpenCL and we are still work with the language a lot – from embedded GPUs and FPGAs to high-end GPUs. Like OpenCL unjustly isn’t associated with clusters full of professional GPUs, we were not associated with CUDA. I can tell many of our customers have found us to build high performance software in CUDA.

Breaking with the past is not easy due to associations that seem to stick. With the name change from StreamComputing to Stream HPC some years ago, we wanted to enforce that break with being “the OpenCL company”. For some time we were much more pragmatic in solving the problems of our customers, which resulted in making software in MPI and CUDA – sometimes an unexpected direction as the customer initially chose OpenCL.

We also started hiring people who only knew CUDA (but expect them to learn OpenCL), as the right algorithm and the right processor is more important. Internships with CUDA, large CUDA-projects, seeking better relations with Nvidia and such – all have been going on for years. And we like it as much as we like OpenCL – both have unique advantages.

So if you have questions about CUDA, don’t be afraid that you hurt us – we’re happy to help you get fast software.

NVIDIA: mobile phones, tablets and HPC (cloud)

If you want to see what is coming up in the market of consumer-technology (PC, mobile and tablet), then NVIDIA can tell you the most. The company is very flexible, and shows time after time it really knows in which markets is currently operates and can enter. I sometimes strongly disagree with their marketing, but watch them closely as they are in the most important markets to define the near future in: PCs, Mobile/Tablet and HPC.
You might think I completely miss interconnects (buses between processors, devices and memory) and memory-technologies as clouds have a large need for high-speed data-transport, but the last 20 years have shown that this is a quite stable developing market based on IP-selling to the hardware-vendors. With the acquisition of Cray’s interconnect technology, we have seen this is serious business for Intel, so things might change indeed. For this article I want to focus on NVIDIA’s choices.

PDFs of Monday 16 April

By exception, another PDF-Monday.

OpenCL vs. OpenMP: A Programmability Debate. The one moment OpenCL and the other mom ent OpenMP produces faster code. From the conclusion: “OpenMP is more productive, while OpenCL is portable for a larger class of devices. Performance-wise, we have found a large variety of ratios between the two solutions, depending on the application, dataset sizes, compilers, and architectures.”

Improving Performance of OpenCL on CPUs. Focusing on how to optimise OpenCL. From the abstract: “First, we present a static analysis and an accompanying optimization to exclude code regions from control-flow to data-flow conversion, which is the commonly used technique to leverage vector instruction sets. Second, we present a novel technique to implement barrier synchronization.”

Variants of Mersenne Twister Suitable for Graphic Processors. Source-code at

Accelerating the FFTD method using SSE and GPUs. “The Finite-Difference Time-Domain (FDTD) method is a computational technique for modelling the behaviour of electromagnetic waves in 3D space”. This is a project-plan, but describes the theories pretty well. Continue reading “PDFs of Monday 16 April”

PDFs of Monday 19 September

Already the fourth PDF-Monday. It takes quite some time, so I might keep it to 10 in the future – but till then enjoy! Not sure which to read? Pick the first one (for the rest there is not order).

Edit: and the last one, follow me on twitter to see the PDFs I’m reading. Reason is that hardly anyone clicked on the links to the PDFs.

I would like if you let others know in the comments which PDF you liked a lot.

Adding Physics to Animated Characters with Oriented Particles (Matthias Müller and Nuttapong Chentanez). Discusses how to accelerate movements of pieces of cloth attached to the bodies. Not time to read? There are nice pictures.

John F. Peddy’s analysis on the GPU market.

Hardware/Software Co-Design. Simple Solution to the Matrix Multiplication Problem using CUDA.

CUDA Based Algorithms for Simulating Cardiac Excitation Waves in a Rabbit Ventricle. Bioinformatics.

Real-time implementation of Bayesian models for multimodal perception using CUDA.

GPU performance prediction using parametrized models (Master-thesis by Andreas Resios)

A Parallel Ray Tracing Architecture Suitable for Application-Specific Hardware and GPGPU Implementations (Alexandre S. Nery, Nadia Nedjah, Felipe M.G. Franca, Lech Jozwiak)

Rapid Geocoding of Satellite SAR Images with Refined RPC Model. An ESA-presentation by Lu Zhang, Timo Balz and Mingsheng Liao.

A Parallel Algorithm for Flight Route Planning with CUDA (Master-thesis by Seçkîn Sanci). About the travelling salesman problem and much more.

Color-based High-Speed Recognition of Prints on Extruded Materials. Product-presentation on how to OCR printed text on cables.

Supplementary File of Sequence Homology Search using Fine-Grained Cycle Sharing of Idle GPUs (Fumihiko Ino, Yuma Munekawa, and Kenichi Hagihara). They sped up the BOINC-system (Folding@Home). Bit vague what they want to tell, but maybe you find it interesting.

Parallel Position Weight Matrices Algorithms (Mathieu Giraud, Jean-Stéphane Varré). Bioinformatics, DNA.

GPU-based High Performance Wave Propagation Simulation of Ischemia in Anatomically Detailed Ventricle (Lei Zhang, Changqing Gai, Kuanquan Wang, Weigang Lu, Wangmeng Zuo). Computation in medicine. Ischemia is a restriction in blood supply, generally due to factors in the blood vessels, with resultant damage or dysfunction of tissue

Per-Face Texture Mapping for Realtime Rendering. A Siggraph2011 presentation by Disney and NVidia.

Introduction to Parallel Computing. The CUDA 101 by Victor Eijkhout of University of Texas.

Optimization on the Power Efficiency of GPU and Multicore Processing Element for SIMD Computing. Presentation on what you find out when putting the volt-meter directly on the GPU.

NUDA: Programming Graphics Processors with Extensible Languages. Presentation on NUDA to write less code for GPGPU.

Qt FRAMEWORK: An introduction to a cross platform application and user interface framework. Presentation on the Qt-platform – which has great #OpenCL-support.

Data Assimilation on future computer architectures. The problems projected for 2020.

Current Status of Standards for Augmented Reality (Christine Perey1, Timo Engelke and Carl Reed). not much to do with OpenCL, but tells an interesting purpose for it.

Parallel Computations of Vortex Core Structures in Superconductors (Master-thesis by Niclas E. Wennerdal).

Program the SAME Here and Over There: Data Parallel Programming Models and Intel Many Integrated Core Architecture. Presentation on how to program the Intel MIC.

Large-Scale Chemical Informatics on GPUs (Imran S. Haque, Vijay S. Pande). Book-chapter on the design and optimization of GPU implementations of two popular chemical similarity techniques: Gaussian shape overlay (GSO) and LINGO.

WebGL, WebCL and Beyond! A presentation by Neil Trevett of NVidia/Khronos.

Biomanycores, open-source parallel code for many-core bioinformatics (Mathieu Giraud, Stéphane Janot, Jean-Frédéric Berthelot, Charles Delte, Laetitia Jourdan , Dominique Lavenier , Hélène Touzet, Jean-Stéphane Varré). A short description on the project

PDFs of Monday 12 September

As it got more popular that I shared my readings, I decided to put them on my site. I focus on everything that uses vector-processing (GPUs, heterogeneous computing, CUDA, OpenCL, GPGPU, etc). Did I miss something or you have a story you want to share? Contact me or comment on this article. If you tell others about these projects you discovered here, I would appreciate you mention my website or my twitter @StreamHPC.

The research-papers have their authors mentions; the other links can be presentations or overviews of (mostly) products. I have read all, except the long PhD-theses (which are on my non-ad-hoc reading-list) – drop me any question you have.

Bullet Physics, Autodesk style. AMD and Autodesk on integrating Bullet Physics engine into Maya.

MERCUDA: Real-time GPU-based marine scene simulation. OpenCL has enabled more realistic sea and sky simulation for this product, see page 7.

J.P.Morgan: Using Graphic Processing Units (GPUs) in Pricing and Risk. Two pages describing OpenCL/CUDA can give 10 to 100 times speedup over conventional methods.

Parallelization of the Generalized Hough Transform on GPU (Juan Gómez-Luna1a, José María González-Linaresb, José Ignacio Benavidesa, Emilio L. Zapatab and Nicolás Guil). Describing two parallel methods for the Fast Generalized Hough Transform (Fast GHT) using GPUs, implemented in CUDA. It studies how load balancing and occupancy impact the performance of an application on a GPU. Interesting article as it shows that you can choose in which limits you bump into.

Performance Characterization and Optimization of Atomic Operations on AMD GPUs (Marwa Elteir, Heshan Lin and Wu-chun Feng). Measurement of the impact of using atomic operations on AMD GPUs. It seems that even mentioning ‘atomic’ puts the kernel in atomic mode and has major influence on the performance. They also come up with a solution: software-based atomic operation. Work in progress.

On the Efficacy of a Fused CPU+GPU Processor (or APU) for Parallel Computing (Mayank Daga, Ashwin M. Aji, and Wu-chun Feng). Another one from Virginia Tech, this time on AMD’s APUs. This article measures its performance via a set of micro-benchmarks (e.g., PCIe data transfer), kernel benchmarks (e.g., reduction), and actual applications (e.g., molecular dynamics). Very interesting to see in which cases discrete GPUs have a disadvantage even with more muscle power.

A New Approach to rCUDA (José Duato, Antonio J. Peña, Federico Silla1, Juan C. Fernández, Rafael Mayo, and Enrique S. Quintana-Ort). On (remote) execution of CUDA-software within VMs. Interesting if you want powerful machines in your company to delegate heavy work to, or are interested in clouds.

Parallel Smoothers for Matrix-based Multigrid Methods on Unstructured Meshes Using Multicore CPUs and GPUs (Vincent Heuveline, Dimitar Lukarski, Nico Trost and Jan-Philipp Weiss). Different methods around 8 multi-colored Gauß-Seidel type smoothers using OpenMP and GPUs. Also some words on scalability!

Visualization assisted by parallel processing (B. Lange, H. Rey, X. Vasques, W. Puech and N. Rodriguez). How to use GPGPU for visualising big data. An important factor of real-time data-processing is that people get more insight in the matter. As an example they use temperatures in a server-room. As I see more often now, they benchmark CPU, GPU and hybrids.

A New Tool for Classification of Satellite Images Available from Google Maps: Efficient Implementation in Graphics Processing Units (Sergio Bernabéa and Antonio Plaza). 30 times speed-up with a new parallel implementation of the k-means unsupervised clustering algorithm in CUDA. Ity is used for classification of satellite images.

TAU performance System. Product-presentation of TAU which does, among other things, parallel profiling and tracing. Support for CUDA and OpenCL. Extensive collection of tools, so worth to spend time on. An paper released in March describes TAU and compares it with two other performance measurement systems: PAPI and VamirTrace.

An Experimental Approach to Performance Measurement of Heterogeneous Parallel Applications using CUDA (Allen D. Malony, Scott Biersdorff, Wyatt Spear and Shangkar Mayanglambam). Using a TAU-based (see above) tool TAUcuda this paper describes where to focus on when optimising heterogeneous systems.

Speeding up the MATLAB complex networks package using graphic processors (Zhang Bai-Da, Wu Jun-Jie, Tang Yu-Hua and Li Xin). Free registration required. Their conclusion: “In a word, the combination of GPU hardware and MATLAB software with Jacket Toolbox enables high-performance solutions in normal server”. Another PDF I found was: Parallel High Performance Computing with emphasis on Jacket based computing.

Profile-driven Parallelisation of Sequential Programs (Georgios Tournavitis). PhD-thesis on a new approach for extracting and exploiting multiple forms of coarse-grain parallelism from sequential applications written in C.

OpenCL, Heterogeneous Computing, and the CPU. Presentation by Tim Mattson of Intel on how to use OpenCL with the vector-extensions of Intel-processors.

MMU Simulation in Hardware Simulator Based-on State Transition Models (Zhang Xiuping, Yang Guowu and Zheng Desheng). It seems a bit off-chart to have a paper on the Memory Management Unit of a ARM, but as the ARM-processor gets more important some insights on its memory-system is important.

Multi-Cluster Performance Impact on the Multiple-Job Co-Allocation Scheduling (Héctor Blanco, Eloi Gabaldón, Fernando Guirado and Josep Lluí Lérida). This research-group has developed a scheduling-technique, and in this paper they discuss in which situations theirs works better than existing techniques.

Convey Computers: Putting Personality Into High Performance Computing. Product-presentation. They combine X86-CPUs with pre-programmed FPGAs to get high though-put. In short: if you make heavy usage of the provided algorithms, then this might be an alternative to GPGPU.

High-Performance and High-Throughput Computing. What it means for you and your research. Presentation by Philip Chan of Monash University. Though the target-group is their own university, it gives nice insights on how it goes around on other universities and research-groups. HPC is getting cheaper and accepted in more and more types of research.

Bull: Porting seismic software to the GPU. Presentation for oil-companies on finding new oil-fields. These seismic calculations are quite computation-intensive and therefore portable HPC is needed. Know StreamHPC is also assisting in porting such code to GPUs.

Dymaxion: Optimizing Memory Access Patterns for Heterogeneous Systems (Shuai Che, Jeremy W. Sheaffer and Kevin Skadron). This piece of software allows CUDA-programmers to optimize memory mappings to improve the efficiency of memory accesses on heterogeneous platforms.

Real-time volumetric shadows for dynamic rendering (MsC-thesis of Alexandru Teodor V.L. Voicu). Self-shadowing using the Opacity Shadow Maps algorithm is not fit for real-time processing. This thesis discusses Bounding Opacity Maps, a novel method to overcome this problem. Including code at the end, which you can download here.

Accelerating Foreign-Key Joins using Asymmetric Memory Channels (Holger Pirk, Stefan Manegold and Martin Kersten). Shows how to accelerate Foreign-Key Joins by executing the random table lookups on the GPU’s VRAM while sequentially streaming the Foreign-Key-Index through the PCI-E Bus. Very interesting on how to make clever usage of I/O-bounds.

Come back next Monday for more interesting research papers and product presentations. If you have questions, don’t hesitate to contact StreamHPC.

PDFs of Monday 5 September

Live from le Centre Pompidou in Paris: Monday PDF-day. I have never been inside the building, but it is a large public library where people are queueing to get in – no end to the knowledge-economy in Paris. A great place to read some interesting articles on the subjects I like.

CUDA-accelerated genetic feedforward-ANN training for data mining (Catalin Patulea, Robert Peace and James Green). Since I have some background on Neural Networks, I really liked this article.

Self-proclaimed State-of-the-art in Heterogeneous Computing (Andre R. Brodtkorb a , Christopher Dyken, Trond R. Hagen, Jon M. Hjelmervik, and Olaf O. Storaasli). It is from 2010, but just got thrown on the net. I think it is a must-read on Cell, GPU and FPGA architectures, even though (as also remarked by others) Cell is not so state-of-the-art any more.

OpenCL: A Parallel Programming Standard for Heterogeneous Computing Systems (John E. Stone, David Gohara, and Guochun Shi). A basic and clear introduction to my favourite parallel programming language.

Research proposal: Heterogeneity and Reconfigurability as Key Enablers for Energy Efficient Computing. About increasing energy efficiency with GPUs and FPGAs.

Design and Performance of the OP2 Library for Unstructured Mesh Applications. CoreGRID presentation/workshop on OP2, an open-source parallel library for unstructured grid computations.

Design Exploration of Quadrature Methods in Option Pricing (Anson H. T. Tse, David Thomas, and Wayne Luk). Accelerating specific option pricing with CUDA. Conclusion: FPGA has the least Watt per FLOPS, CUDA is the fastest, and CPU is the big loser in this comparison. Must be mentioned that GPUs are easier to program than FPGAs.

Technologies for the future HPC systems. Presentation on how HPC company Bull sees the (near) future.

Accelerating Protein Sequence Search in a Heterogeneous Computing System (Shucai Xiao, Heshan Lin, and Wu-chun Feng). Accelerating the Basic Local Alignment Search Tool (BLAST) on GPUs.

PTask: Operating System Abstractions To Manage GPUs as Compute Devices (Christopher J. Rossbach, Jon Currey, Mark Silberstein, Baishakhi Ray, and Emmett Witchel). MS research on how to abstract GPUs as compute devices. Implemented on Windows 7 and Linux, but code is not available.

PhD thesis by Celina Berg: Building a Foundation for the Future of Software Practices within the Multi-Core Domain. It is about a Rupture-model described at Ch.2.2.2 (PDF-page 59). [total 205 pages].

Workload Balancing on Heterogeneous Systems: A Case Study of Sparse Grid Interpolation (Alin Murarasu, Josef Weidendorfer, and Arndt Bodes). To my opinion a very important subject as this can help automate much-needed “hardware-fitting”.

Fraunhofer: Efficient AMG on Heterogeneous Systems (Jiri Kraus and Malte Förster). AMG stands for Algebraic MultiGrid method. Paper includes OpenCL and CUDA benchmarks for NVidia hardware.

Enabling Traceability in MDE to Improve Performance of GPU Applications (Antonio Wendell de O. Rodrigues, Vincent Aranega, Anne Etien, Frédéric Guyomarc’h, Jean-Luc Dekeyser). Ongoing work on OpenCL code generation from UML (Model Driven Design). [34 pag PDF]

GPU-Accelerated DNA Distance Matrix Computation (Zhi Ying, Xinhua Lin, Simon Chong-Wee See and Minglu Li). DNA sequences distance computation: [PDF] #OpenCL #GPGPU #Biology

And while browsing around for PDFs I found the following interesting links:

  • Say bye to Von Neumann. Or how IBM’s Cognitive Computer Works.
  • Workshop on HPC and Free Software. 5-7 October 2011, Ourense, Spain. Info via
  • Basic CUDA course, 10 October, Delft, Netherlands, €200,-.
  • Par4All: automatic parallelizing and optimizing compiler for C and Fortran sequential programs.
  • LAMA: Library for Accelerated Math Applications for C/C++.

PDFs of Monday 29 August

This is the first PDF-Monday. It started as I used Mondays to read up on what happens around OpenCL and I like to share with you. It is a selection of what I find (somewhat) interesting – don’t hesitate to contact me on anything you want to know about accelerated software.

Parallel Programming Models for Real-Time Graphics. A presentation by Aaron Lefohn of Intel. Why a mix of data-, task-, and pipeline-parallel programming works better using hybrid computing (specifically Intel processors with the latest AVX and SSE extensions) than using GPGPU.

The Practical Reality of Heterogeneous Super Computing. A presentation of Rob Farber of NVidia on why discrete GPUs has a great future even if heterogeneous processors hit the market. Nice insights, as you can expect from the author of the latest CUDA-book.

Scalable Simulation of 3D Wave Propagation in Semi-Infinite Domains Using the Finite Difference Method (Thales Luis Rodrigues Sabino, Marcelo Zamith, Diego Brandâo, Anselmo Montenegro, Esteban Clua, Maurício Kischinhevksy, Regina C.P. Leal-Toledo, Otton T. Silveira Filho, André Bulcâo). GPU based cluster environment for the development of scalable solvers for a 3D wave propagation problem with finite difference methods. Focuses on scattering sound-waves for finding oil-fields.

Parallel Programming Concepts – GPU Computing (Frank Feinbube) A nice introduction to CUDA and OpenCL. They missed task-parallel programming on hybrid systems with OpenCL though.

Proposal for High Data Rate Processing and Analysis Initiative (HDRI). Interesting if you want to see a physics project where they did not have decided yet to use GPGPU or a CPU-cluster.

Physis: An Implicitly Parallel Programming Model for Stencil Computations on Large-Scale GPU-Accelerated Supercomputers (Naoya Maruyama, Tatsuo Nomura, Kento Sato and Satoshi Matsuoka). A collection of macros for GPGPU, tested on TSUBAME2.

OpenCL vs CUDA Misconceptions

Translation available: Russian/Русский. (Let us know if you have translated this article too… And thank you!)

Last year I explained the main differences between CUDA and OpenCL. Now I want to get some old (and partly) false stories around CUDA-vs-OpenCL out of this world. While it has been claimed too often that one technique is just better, it should be also said that CUDA is better in some aspects, whereas OpenCL is better in others.

Why did I write this article? I think NVIDIA is visionary in both technology and marketing. But as I’ve written before, the potential market for dedicated graphics cards is shrinking and therefore forecasting the end of CUDA on desktop. Not having this discussion opens the door for closed standards and delaying innovation, which can happen on top of OpenCL. The sooner people & companies start choosing for a standard that gives equal competitive advantages, the more we can expect from the upcoming hardware.

Let’s stand by what we have learnt at school when gathering information sources, don’t put all your eggs in one basket! Gather as many sources and references as possible. Please also read articles which claim (and underpin!) why CUDA has a more promising future than OpenCL. If you can, post comments with links to articles you think others should read too. We appreciate contributions!

Also found that Google Insights agrees with what I constructed manually.

Continue reading “OpenCL vs CUDA Misconceptions”

Benchmarks Q1 2011

February Benchmark Month. The idea is that you do at least one of the following benchmarks and put the results on the Khronos Forum. If you encounter any technical problems or you think the benchmark favours a certain brand, discuss it below this post. If I missed a benchmark, please put a comment under this post.

Since OpenCL works on all kinds of hardware, we can find out which is the fastest: Intel, AMD or NVIDIA.I don’t think all benchmarks are fit for IBM’s hardware, but I hope to see results of some IBM Cells too. If all goes well, I’ll show the first results of the fastest cards posted in April . Know that if the numbers are off too much, I might want to see further proof.

Happy benchmarking!

Continue reading “Benchmarks Q1 2011”

Windows on ARM

In 2010 Microsoft got interested in ARM, because of low-power solutions for server-parks. ARM tried to lobby for years to convince Microsoft to port Windows to their architecture and now the result is there. Let’s not look to the past, why they did not do it earlier and depended completely on Intel, AMD/ATI and NVIDIA. NB: This article is a personal opinion, to open up the conversation about Windows plus OpenCL.

While Google and Apple have taken their share on the ARM-OS market, Microsoft wants some too. A wise choice, but again late. We’ve seen how the Windows-PC market was targeted first from the cloud (run services in the browser on any platform) and Apple’s user-friendly eye-candy (A personal computer had to be distinguished from a dull working-machine), then from the smartphones and tablets (many users want e-mail and a browser, not sit behind their desk). MS’s responses were Azure (Cloud, Q1 2010), Windows 7 (OS with slick user-interface, Q3 2009), Windows Phone 7 (Smartphones, Q4 2010) and now Windows 8 (OS for X86 PCs and ARM tablets, 2012 or later).

Windows 8 for ARM will be made with assistance from NVIDIA, Qualcomm and Texas Instruments, according to their press-release [1]. They even demonstrated a beta of Windows 8 running Microsoft Office on ARM-hardware, so it is not just a promise.

How can Microsoft support this new platform and (for StreamHPC more important) what will the consequences be for OpenCL, CUDA and DirectCompute.

Continue reading “Windows on ARM”

NVIDIA’s answer to SandyBridge and Fusion

Intel has Sandy Bridge, AMD has Fusion, now NVIDIA has a combination of CPU and GPU too: Project Denver. The only difference is that it is not X86-based, but an ARM-architecture. And most-probable the most powerful ARM-GPU of 2011.

For years there were ARM-based Systems-on-a-chip: a CPU and a GPU combined (see list below). On the X86-platform the “integrated GPU” was on the motherboard, and since this year now both AMD/ATI and Intel hit this “new market”.The big advantage is that it’s cheaper to produce, is more powerful per Watt (in total) and has good acceleration-potential. NVIDIA does not have X86-chips and would have been the big loser of 2011; they did everything to reinvent themselves: 3D was reintroduced, CUDA was actively developed and pushed (free libraries and tools, university-programs, many books and trainings, Tesla, etc), a mobile Tegra graphics solution [1] (see image at the right), and all existing products got extra backing from the marketing-department. A great time for researchers who needed to get free products in exchange of naming NVIDIA in their research-reports.

NVIDIA chose for ARM; interesting for who is watching the CUDA-vs-OpenCL battle, since CUDA was for GPUs of NVIDIA on X86 and ARM was solely for OpenCL. Period. In the contrary to their other ARM-based chips, this new chip probably won’t be in smartphones (yet); it targets systems that need more GPU-power like CUDA and games.

In a few days the article about Windows-on-ARM is to be released, which completes this article.

Continue reading “NVIDIA’s answer to SandyBridge and Fusion”

DirectCompute’s unpopularity

In the world of GPGPU we have currently 4 players: Khronos OpenCL, NVIDIA CUDA, Microsoft DirectCompute and PathScal ENZO. You probably know CUDA and OpenCL already (or start reading more articles from this blog). ENZO is a 64bit-compiler which serves a small niche-market, and DirectCompute is built on top of CUDA/OpenCL or at least uses the same drivers.

Edit 2011-01-03: I was contacted by Pathscale about my conclusions about ENZO. The reason why not much is out there is that they’re still in closed alpha. Expect more to hear from them about ENZO somewhere in the coming 3 months.

A while ago there was an article introducing OpenCL by David Kanter who claimed on page 4 that DirectCompute will win from CUDA. I quote:

Judging by history though, OpenCL and DirectCompute will eventually come to dominate the landscape, just as OpenGL and DirectX became the standards for graphics.

I twittered that I totally disagreed with him and in this article I will explain why I think that.

Continue reading “DirectCompute’s unpopularity”

Learning both OpenCL and CUDA

Be sure to read Taking on OpenCL where I’ve put my latest insights – also for CUDA.

The two¹ “camps” OpenCL and CUDA both claim you should first learn their language first, after which the other would be easy to learn. I’m from the OpenCL-camp, so I say you should learn OpenCL first, but with a strong emphasis on hardware-architecture understanding. If I had chosen for CUDA I would have said the opposite, so in other words it does not matter which you do first. But psychology tells us that you probably like the first language more since there is where you discovered the magic; also most people do not like to learn a second language which is much alike and does not add a real difference. Most programmers just want to get the job done and both camps know that. Be aware of that.

NVIDIA is very good in marketing their products, AMD has – to say it modest – a lower budget for GPGPU-marketing. As a programmer you should be aware of this difference.

The possibilities of OpenCL are larger than those of CUDA, because of task-parallel programming and support for far more different architectures. At the other side CUDA is much more user-friendly and has a lot of convenience built-in.

Continue reading “Learning both OpenCL and CUDA”

Difference between CUDA and OpenCL 2010


Most GPGPU-enthusiasts have heard of both OpenCL and CUDA. While there are more solutions, these have the most potential. Both techniques are very comparable like a BMW and a Mercedes, but there are some differences. Since the technologies will evolve, we’ll take a look at the differences again next year. We’ve discussed this difference in a with a focus on marketing earlier this year.

Disclaimer: we have a strong focus on OpenCL (but actually for reasons explained in this article).


If you have seen kernels of OpenCL and CUDA, you see the biggest difference might be the prefix “cl_” or the prefix “cu_”, but there is also a difference in terminology.

Matt Harvey (developer of Cuda2OpenCL-translator Swan) has summed up the differences in a presentation “Experiences porting from CUDA to OpenCL” (PDF):

CUDA term OpenCL term
GPU Device
Multiprocessor Compute Unit
Scalar core Processing element
Global memory Global memory
Shared (per-block) memory Local memory
Local memory (automatic, or local) Private memory
kernel program
block work-group
thread work item

As far as I know, the kernel-program is also called a kernel in OpenCL. Personally I like Cuda’s terms “thread” and “per-block memory” more. It is very clear CUDA targets the GPU only, while in OpenCL it an be any device.

Edit 2011-01-15: In a talk by Sami Rosendahl the differences are also discussed.


We would like to present you a benchmark between OpenCL and CUDA with full comparison, but we don’t have enough hardware in-house to do a full benchmark. Below information is what we’ve found on the net and a little bit based on our own experience.

On NVidia hardware, OpenCL is up to 10% slower (see Matt Harvey’s presentation); this is mainly because OpenCL is implemented on top of CUDA-architecture (this shouldn’t be a reason, but to say NVidia has put more energy in CUDA is just a wild guess also). On ATI 4000-series OpenCL is just slow, but gives very comparable to NVidia if compared to the 5000-series. The specialised streaming processors NVidia’s Tesla and AMD’s FireStream really bite each other, while the Playstation 3 unbelievably still wins on some tasks.

The architecture AMD/ATI-hardware is very different from NVidia’s and that’s why a kernel written with a specific brand or GPU in mind just performs better than a version which is not optimised. So if you do a benchmark, it really depends on which kernels you use for it. To be more precise: any benchmark can be written in favour of a specific architecture. Fine-tuning the software to work a maximum speed in current and future(!) hardware for different kinds of datasets is (still) a specialised task for that reason. This is also one of the current problems of GPGPU, but kernel-optimisers will get better.

If you like pictures, Hugh Merz comes to the rescue, who compared CUDA-FFT against FFTW (“the fastest FFT in the West”). The page is offline now, but you it was clear that the data-transfer from and to the GPU is a huge bottleneck and Hugh Merz was rather sceptical about GPU-computing in 2007. He extended his benchmark with the PS3 and a Tesla-s1070 and now you see bigger differences. Since CPUs go multi-multi-core, you cannot tell how big this gap will be in the future; but you can tell the gap will be bigger and CPUs will more and more be programmed like GPUs (massively parallel).

What we learn from this is 1) that different devices will improve if the demands are more clear, and 2) that it will be all about specialisation, since different manufacturers will hear different demands. The latest GPUs from AMD works much better with OpenCL, the next might beat all others in a many or only specific areas in 2011 – who knows? IBM’s Cell-processor is expected to enter the ring outside the home-brew PS3 render-farms, but with what specialised product? NVidia wants to enter high in the HPC-world, and they might even win it. ARM is developing multiple-core CPUs, but will it support OpenCL for a better FLOP/Watt than competitors?

It’s all about the choices manufacturers make, which way CUDA en OpenCL will develop.

Homogeneous vs Heterogeneous

For us the most important reason to have chosen for OpenCL, even if CUDA is more mature. While CUDA only targets NVidia’s GPUs (homogeneous), OpenCL can target any digital device that has an input and an output (very heterogeneous). AMD/ATI and Intel are both on the path of making architectures that are heterogeneous; just like Systems-on-a-Chip (SoCs) based on an ARM-architecture. Watch for our upcoming article about ARM & SoCs.

While I was searching for more information about this difference, I came across a blog-item by RogueWave, which claims something different. I think they switched Intel’s architectures with NVidia’s or he knew things were going to change. In the near future could bring us an x86-chip from NVidia. This will change a lot in the field, so more about this later. They already have an ARM-chip in their Tegra mobile processor, so NVidia/CUDA still has some big bullets.

Missing language-features

Like Java and .NET are very comparable, developers from both side know very well that their favourite feature is missing at the other camp. Most time such a feature is an external library, just built in. Or is it taste? Or even a stack of soapboxes?

OpenCL has:

  • Task-parallel execution mode (to be used on CPUs) – not needed on NVidia’s GPUs.

CUDA has unique features too:

  • FFT library – so in OpenCL you need to have your own kernels for it.
  • Atomic operations – which make double-write threads easier to implement.
  • Hardware texture interpolation – OpenCL has to fall back to a larger kernel or OpenGL.
  • Templating – in openCL you have to create new kernels for every data-type.

In short CUDA certainly has made a lot of things just easier for the developer, but OpenCL has its potential in support for more than just GPUs. All differences are based on this difference in focus-area.

I’m pretty sure this list is not complete at all, and only explains the type of differences. So please come to the LinkedIn GPGPU Users Group to discuss this.

Last words


As it is done with more shared standards, there is no win and no gain to promote it. If you promote it, a lot of companies thank you, but the Rreturn-on-Investments is lower than when you have your own standard. So OpenCL is just used-as-it-is-available, while CUDA is highly promoted; for that reason more people invest in partnerships with NVidia to use CUDA instead of non-profit organisation Khronos. And eventually CUDA-drivers can be ported to IBM’s Cell-processors or to ARM, since it is very comparable to OpenCL. It really depends on the profit NVidia will make with such deals, so who can tell what will happen.

We still think OpenCL will win eventually on consumer-markets (desktop and mobile) because of support for more devices, but CUDA will stay a big player in professional and scientific markets because of the legacy software they are currently building up and the more friendly development-support. We hope they will both exist and help each other push forward, just like OpenGL vs DirectX, nVidia vs ATI, Europe vs the USA vs Asia, etc. Time will tell what features will eventually end up in each technology.

Update August 2012: due to higher demand StreamHPC is explicitly offering CUDA to OpenCL porting.

Does GPGPU have a bright future?

This post has a focus towards programmers. The main question “should I invest in learning CUDA/OpenCL?”

Using the video-processor for parallel processing is actually possible since beginning 2006; you just had to know how to use the OpenGL Shader Language. Not long after that (end 2006) CUDA was introduced. A lot has happened after that, which resulted in the introduction of OpenCL in fall 2008. But actually the acceptance of OpenCL is pretty low. Many companies which do use it, want to have it as their own advantage and don’t tell the competition they just saved hundreds of thousands of Euros/Dollars because they could replace their compute-cluster with a single computer which cost them €10 000,- and a rewrite of the calculation-core of their software. Has it become a secret weapon?

This year a lot of effort will be put to integrate OpenCL within the existing programming languages (without all the thousands of tweak-options visible). Think about wizards around pre-built kernels and libraries. Next year everything will be around kernel-development (kernels are the programs which do the actual calculations on the graphics processor). The year after that, the peak is over and nobody knows it is built in their OS or programming-language. It’s just like current programmers use security-protocols, but don’t know what it actually is.

If I want to slide to the next page on modern mobile phones, I just make a call to a slide-function. A lot is happening when the function is called, such building up the next page in a separate part of memory, calling the GPU-functions to show the slide, possibly unloading the previous page. The same is with OpenCL; I want to calculate a FFT with specified precision and I don’t want to care on which device the calculation is done. The advantage of building blocks (like LEGO) is that we keeps the focus of development on the end-target, while we can tweak it later (if the customer has paid for this extra time). What’s a bright future if nobody knows it?

Has it become a secret weapon?

Yes and no. Companies want to brass about their achievements, but don’t want the competitors to go the same way and don’t want their customers to demand lower prices. AMD and NVidia are pushing OpenCL/CUDA, so it won’t stop growing in the market, but actually this pushing is the biggest growth in the market. NVidia does a good job with marketing their CUDA-platform.

What’s a bright future if nobody knows it?

Everything that has market-wide acceptation has a bright future. It might be replaced by a successor, but acceptance is the key. With acceptance there always will be a demand for (specialised) kernels to be integrated in building blocks.

We also have the new processors with 32+ cores, which actually need to be used; you know the problem with dual-core “support”.

Also the mobile market is growing rapidly. Once that is opened for OpenCL, there will be a huge growth in demand for accelerated software.

My advise: if high performance is very important for your current or future tasks, invest in learning how to write kernels (CUDA or OpenCL, whatever your favourite is). Use wrapper-libraries which make it easy for you, because once you’ve learned how to use the OpenCL-calls they are completely integrated in your favourite programming language.

nVidia’s CUDA vs OpenCL Marketing

Please read this article about Microsoft and OpenCL, before reading on. The requested benchmarking is done by the people of Unigine have some results on differences between the three big GPGPU-APIs: part I, part II and part III with some typical results.
The following read is not about the technical differences of the 3 APIs, but more about the reason behind why alternate APIs are being kept maintained while OpenCL does the trick. Please comment, if you think OpenCL doesn’t do the trick

As was described in the article, it is important for big companies (such as Microsoft and nVidia) to defend their market-share. This protection is not provided through an open standard like OpenCL. As it went with OpenGL – which was sort of replaced by DirectX to gain market-share for Windows – now nVidia does the same with CUDA. First we will sum up the many ways nVidia markets Cuda and then we discuss the situation.

The way nVidia wanted to play the game, was soon very clear: it wanted to market CUDA to be seen as the better alternative for OpenCL. And it is doing a very good job.

The best way to get better acceptance is giving away free information, such as articles and courses. See Cuda’s university courses as an example. Also sponsoring helps a lot, so the first main-stream oriented books about GPGPU discussed Cuda in the first place and interesting conferences were always supported by Cuda’s owner. Furthermore loads of money is put into an expensive webpage with very complete information about GPGPU there is to find. nVidia does give a choice, by also having implemented OpenCL in its drivers – it does just not have big pages on how-to-learn-OpenCL.

AMD – having spent their money on buying ATI, could not put this much effort in the “war” and had to go for OpenCL. You have to know, AMD has faster graphics-cards for lower prices than nVidia at the moment; so based on that, they could become the winners on GPGPU (if that was to only thing to do). Intel saw the light too late and is even postponing their High-end GPU, the Larrabee. The only help for them is that Apple demands to have OpenCL on nVidia’s drivers – but for how long? Apple does not want strict dependency on nVidia, since it i.e. also has a mobile market. But what if all Apple-developers create their applications on CUDA?

Most developers – helped by the money&support of nVidia – see that there is just little difference between Cuda and OpenCL and in case of a changing market they could translate their programs from one to the other. For now a demand to have a high-end videocard of nVidia can be rectified, a card which actually many people have or easily could buy within the budget of their current project. The difference between Cuda and OpenCL is comparable with C# and Java – the corporate tactics are also the same. Possibly nVidia will have better driver-support for Cuda than OpenCL and since Cuda does not work on AMD-cards, the conclusion is that Cuda is faster. There can then be a situation that AMD and Intel have to buy Cuda-patents, since OpenCL does not have the support.

We hope that OpenCL will stay the main-stream GPGPU-API, so the battle will be on hardware/drivers and support of higher-level programming-languages. We do really appreciate what nVidia already has done for the GPGPU-industry, but we hope they will solely embrace OpenCL for the sake of the long-term market-development.

What we left out of this discussion is Microsoft’s DirectCompute. It will be used by game-developers in the Windows-platform, which just need physics-calculations. When discussing the game-industry, we will tell more about Microsoft’s DirectX-extension.
Also come back and read our upcoming article about FPGAs + OpenCL, a combination from which we expect a lot.