Serving the Quantitative Finance Community

• 1
• 2
• 3
• 4
• 5
• 10

abdelali
Topic Author
Posts: 9
Joined: September 30th, 2009, 8:35 am

### sample cuda problems in finance

Hi there, I am willing to sharpen my skills in CUDA in the finance domain. I would be very thankful if you could propose some problems that need acceleration in finance apart from Monte Carlo, Finite Differences and Finite Elements. I am more interested in problems that have to do with algorithmic trading or data analysis. thanks a lot.

jambodev
Posts: 80
Joined: September 6th, 2008, 11:07 am

### sample cuda problems in finance

There is a feature request for this, pending, in quantlib if I'm not mistaken.

lballabio
Posts: 983
Joined: January 19th, 2004, 12:34 pm

### sample cuda problems in finance

Except he said "apart from Monte Carlo, Finite Differences and Finite Elements." Those would be the obvious additions to QuantLib.

Cuchulainn
Posts: 64978
Joined: July 16th, 2004, 7:38 am
Location: Drosophila melanogaster
Contact:

### sample cuda problems in finance

Quote I am more interested in problems that have to do with algorithmic trading or data analysis. Question is in how far CUDA is suitable for these work-flow/pipeline applications. Or are you thinking about some kind of SPMD? In general, you have MPMD, yes?
Last edited by Cuchulainn on September 29th, 2009, 10:00 pm, edited 1 time in total.
"Compatibility means deliberately repeating other people's mistakes."
David Wheeler

http://www.datasimfinancial.com
http://www.datasim.nl

abdelali
Topic Author
Posts: 9
Joined: September 30th, 2009, 8:35 am

### sample cuda problems in finance

QuoteThere is a feature request for this, pending, in quantlib if I'm not mistaken.I am considering this one already. thx
Last edited by abdelali on September 29th, 2009, 10:00 pm, edited 1 time in total.

abdelali
Topic Author
Posts: 9
Joined: September 30th, 2009, 8:35 am

### sample cuda problems in finance

QuoteOriginally posted by: CuchulainnQuote I am more interested in problems that have to do with algorithmic trading or data analysis. Question is in how far CUDA is suitable for these work-flow/pipeline applications. Or are you thinking about some kind of SPMD? In general, you have MPMD, yes?Truth is. I don't know exactly. at the end of the day, it will be SPMD to run on CUDA. it is just just thats some MPMD some problems can be reformulated to fit the SPMD paradigm. The goal here is to gather a list of problems that can be accelerated, see which ones can be CUDified (completely or partially), CUDify them and develop some patterns and good practices reference along the way. It is really a pitty that we do not have open high performance financial libraries. Other disciplines do, why don't we ?

abdelali
Topic Author
Posts: 9
Joined: September 30th, 2009, 8:35 am

### sample cuda problems in finance

there are now two more reasons to dig deeper in the GPU spaceNexushttp://www.nvidia.com/object/pr_nexus_093009.htmland Fermihttp://www.nvidia.com/object/fermi_architecture.htmlwhat do you guys think ?

Cuchulainn
Posts: 64978
Joined: July 16th, 2004, 7:38 am
Location: Drosophila melanogaster
Contact:

### sample cuda problems in finance

what do you guys think ? My 2 cents; those vendors who adopt open standards (C++, OpenCL, IEEE) will win. At least that's how the history of s/w products have evolved.In the 80's it was the software that sold the hardware. Think about  CAD systems running in VAX, UNIX and Pr1me boxes. Early pioneers tended to fall by the wayside. Anyone rememberTaligentOccamOS/2 (except DCFC )NextUltrix...
Last edited by Cuchulainn on September 30th, 2009, 10:00 pm, edited 1 time in total.
"Compatibility means deliberately repeating other people's mistakes."
David Wheeler

http://www.datasimfinancial.com
http://www.datasim.nl

abdelali
Topic Author
Posts: 9
Joined: September 30th, 2009, 8:35 am

### sample cuda problems in finance

I completely agree with you Cuchulainn. On the long run OpenCL or the standard that comes after it will win. However this does not affect my original question. The patterns will be the same whether we use CUDA or OpenCL. so the most important thing for me now is to put together a list of problems in financial industry that are currently slow/not fast enough. To do that, I need your (the community's) input as my expertise is limited to pricing and scenario generation. what do you think ?

DominicConnor
Posts: 11684
Joined: July 14th, 2002, 3:00 am

### sample cuda problems in finance

I think you're probabily right, but history is more ambiguous.Partly because we're talking here about the relationship between programming language and architecture.The Algol family (Pascal, C++, C, Java, C#) so dominate programming that it is easy to forget there is any other way.However they assume a classical von Neuman architecture, unlike (say) the functional languages that followed Lisp.Algols are not ideal multi tasking, which is of course ironic since almost all mutlitasking is controlled by the C/C++ family which either does it natively or manages it for the Java/C# class of languages.CUDA is a mutant C, and we hear that it will move towards C++, which is good, but again not wholly ideal.Another factor is an irony, that the tools for developing languages make it orders of magnitude easier to create or implement one that it was when C was created,but...But it has become stupidly hard to make money from languages, most are either given away for free (GCC, most Java, Perl, Python, and of course CUDA), or sold at a loss like VC++, MS now actually gives away an excellently standard C++, together with a professional development environment.It was not unknown as late as the 80s for a compiler to be so expensive you needed board level approval to buy it.Sun has never made money from Java, unless you make the most heroic and unrealistic assumptions, (ie you read Sun annual reports).In fact the whole developer landscape is in effect a loss leader for other activities and hobbies, including and especially GPUs.This means a tension between what is good for the product and what is good for the people paying the bills.MS tools are hampered by the desire to suck people in, but not let them out, IBM tools suffer from the need to find employment for arts graduates, open source to pursue social agendas, and GPUs to sell hardware.That means that the h/w vendors will try harder on their proprietary shit than the common tools, and inevitably we will see proprietary "extensions".
Last edited by DominicConnor on October 1st, 2009, 10:00 pm, edited 1 time in total.

abdelali
Topic Author
Posts: 9
Joined: September 30th, 2009, 8:35 am

### sample cuda problems in finance

again!! I really do not care which one wins. the future is a cone not a line. So as long as we stay inside the cone, we are fine. developing patterns for GPU programming within the financial industry is a safe bet (in my humble opinion). So guys please, let's refocus on the original question. What would be interesting topics to work on finance (problems of which the solution can/could be accelerated)?

i386
Posts: 100
Joined: January 31st, 2007, 11:21 am

### sample cuda problems in finance

In the long run, OpenCL or its successor might be the winner but if you look for something that can run at the moment or in 1-2 years, CUDA is nearly the only choice, practically speaking. The API provided by OpenCL is too much focused on the graphical side. It doesn't look like a language as CUDA does.

Cuchulainn
Posts: 64978
Joined: July 16th, 2004, 7:38 am
Location: Drosophila melanogaster
Contact:

### sample cuda problems in finance

QuoteIn the long run, OpenCL or its successor might be the winner And there again, it might not...
"Compatibility means deliberately repeating other people's mistakes."
David Wheeler

http://www.datasimfinancial.com
http://www.datasim.nl

Alan
Posts: 10712
Joined: December 19th, 2001, 4:01 am
Location: California
Contact:

### sample cuda problems in finance

QuoteOriginally posted by: abdelaliHi there, I am willing to sharpen my skills in CUDA in the finance domain. I would be very thankful if you could propose some problems that need acceleration in finance apart from Monte Carlo, Finite Differences and Finite Elements. I am more interested in problems that have to do with algorithmic trading or data analysis. thanks a lot.There are a zillion opportunities here, but most of them require good access to live trading databases and/or goodclean historical data. For example, if you have OPRA and stock feeds, you could develop some nice live updating 'smile'displays -- say a display showing the market smiles of the 100 most active optionable securities. You could overlaythat data with some fitted parametric forms (I mention SABR and Gatheral's SVI fit in a thread in the General forum.)

untler
Posts: 15
Joined: June 18th, 2009, 7:23 pm

### sample cuda problems in finance

QuoteOriginally posted by: AlanQuoteOriginally posted by: abdelaliHi there, I am willing to sharpen my skills in CUDA in the finance domain. I would be very thankful if you could propose some problems that need acceleration in finance apart from Monte Carlo, Finite Differences and Finite Elements. I am more interested in problems that have to do with algorithmic trading or data analysis. thanks a lot.There are a zillion opportunities here, but most of them require good access to live trading databases and/or goodclean historical data. For example, if you have OPRA and stock feeds, you could develop some nice live updating 'smile'displays -- say a display showing the market smiles of the 100 most active optionable securities. You could overlaythat data with some fitted parametric forms (I mention SABR and Gatheral's SVI fit in a thread in the General forum.)i was thinking the same thing - but i have no data.might be interesting to play around with this anyway.