SERVING THE QUANTITATIVE FINANCE COMMUNITY

  • 1
  • 6
  • 7
  • 8
  • 9
  • 10
 
User avatar
Traden4Alpha
Posts: 23951
Joined: September 20th, 2002, 8:30 pm

sample cuda problems in finance

April 28th, 2015, 1:50 pm

QuoteOriginally posted by: outrunCuch trying dual core breakfastThe cereal port looks like the bottleneck in this process. Buffer overflow soon?
 
User avatar
Cuchulainn
Posts: 62395
Joined: July 16th, 2004, 7:38 am
Location: Amsterdam
Contact:

sample cuda problems in finance

April 28th, 2015, 2:26 pm

QuoteOriginally posted by: outrunYou're on a different planet, I'm always surprised about the conflict between the way you act and at the same time trying to promote your business.It's a good job for you that you do not post your real name here. I bet you are scared :D
Last edited by Cuchulainn on April 27th, 2015, 10:00 pm, edited 1 time in total.
 
User avatar
Cuchulainn
Posts: 62395
Joined: July 16th, 2004, 7:38 am
Location: Amsterdam
Contact:

sample cuda problems in finance

April 28th, 2015, 2:38 pm

QuoteOriginally posted by: outrunQuoteOriginally posted by: CuchulainnQuoteOriginally posted by: outrunQuoteOriginally posted by: CuchulainnQuoteOriginally posted by: outrunQuoteOriginally posted by: CuchulainnHere is another approach Quote* Only very old GPUs don't have doubles, that was the first generation.So I can throw that CUDA book you recommended in the bin; it uses float.Yes, you can put it on a blanked outside your home if you can't come up with the obvious solution yourself. I've said it before: I'm an expert in spoon feeding a toddler Brussels sprouts -one of the most difficult tasks on earth-, and you are acting like one.Indeed; communication is not your strong point.You're on a different planet, I'm always surprised about the conflict between the way you act and at the same time trying to promote your business.It's a good job for you that you do not post your real name here. I bet you are scared :DAre you kidding? Everybody knows my name. You're the one that he an "special person" reputation here and think it's important to try to manage perception, but you can't help it. I'm a WYSIWYG guy, maybe a bit intolerant for negative people with inflated egos like you but that's not a big deal.Booh! Write your name on your avatar.Dutch love praising themselves I have noticed, you are sooo direct :D So, nothing
Last edited by Cuchulainn on April 27th, 2015, 10:00 pm, edited 1 time in total.
 
User avatar
mj
Posts: 3449
Joined: December 20th, 2001, 12:32 pm

sample cuda problems in finance

May 23rd, 2015, 3:12 am

sorry to be a bit late to the conversation.I have implemented Monte Carlo pricing of IRD with the LMM on the GPU with least-squares for early exercise features.You can get the code from kooderive.sourceforge.net in both C++ and CUDA. The paper is at http://ssrn.com/abstract=2388415I used a completely different code for CUDA than I had previously used for C++. In essence, I treat data as the central concept and use the code to act on the data. The style is very functional. It did take a lot of work because my previous C++ implementations had been object oriented.
 
User avatar
mj
Posts: 3449
Joined: December 20th, 2001, 12:32 pm

sample cuda problems in finance

May 24th, 2015, 12:23 am

indeed threads in the same half-warp are essentially in a SIMD architecture situation.
 
User avatar
AlexEro
Posts: 203
Joined: November 25th, 2014, 4:27 pm
Location: Ukraine
Contact:

sample cuda problems in finance

May 26th, 2015, 5:09 am

QuoteOriginally posted by: mjindeed threads in the same half-warp are essentially in a SIMD architecture situation.The relation between them is not so simple.SIMT GPU SIMD"Programmable NVIDIA GPUs are very inspiring to hardware geeks, proving that processors with an original, incompatible programming model can become widely used.NVIDIA call their parallel programming model SIMT - "Single Instruction, Multiple Threads". Two other different, but related parallel programming models are SIMD - "Single Instruction, Multiple Data", and SMT - "Simultaneous Multithreading". Each model exploits a different source of parallelism: In SIMD, elements of short vectors are processed in parallel. In SMT, instructions of several threads are run in parallel. SIMT is somewhere in between ? an interesting hybrid between vector processing and hardware threading.My presentation of SIMT is focused on hardware architecture and its implications on the trade-off between flexibility and efficiency. I'll describe how SIMT is different from SIMD and SMT, and why ? what is gained (and lost) through these differences."
 
User avatar
AlexEro
Posts: 203
Joined: November 25th, 2014, 4:27 pm
Location: Ukraine
Contact:

sample cuda problems in finance

May 26th, 2015, 5:19 am

QuoteOriginally posted by: mjsorry to be a bit late to the conversation.I have implemented Monte Carlo pricing of IRD with the LMM on the GPU with least-squares for early exercise features.You can get the code from kooderive.sourceforge.net in both C++ and CUDA. The paper is at http://ssrn.com/abstract=2388415I used a completely different code for CUDA than I had previously used for C++. In essence, I treat data as the central concept and use the code to act on the data. The style is very functional. It did take a lot of work because my previous C++ implementations had been object oriented.Very impressive that You can reach 100X...137X speedup on a so complicated algorithm. While personally I am strongly against OOP and C++, but I think Your article deserves to be expanded to a small book - with detailed project information AND detailed explanation of design decisions during this work.There is a lack of this kind of detailed books on the market.As I understand it took 3 years to complete a project, taking into account that You have got some parts ready in CPU version, right?
 
User avatar
Cuchulainn
Posts: 62395
Joined: July 16th, 2004, 7:38 am
Location: Amsterdam
Contact:

sample cuda problems in finance

May 26th, 2015, 6:37 am

QuoteAs I understand it took 3 years to complete a project, taking into account that You have got some parts ready in CPU version, right?The number of manhours is a good indicator. A project manager might like to know that number. Out of curiosity, a guess: [600, 1000] hours?
Last edited by Cuchulainn on May 25th, 2015, 10:00 pm, edited 1 time in total.
 
User avatar
bojan
Posts: 347
Joined: August 8th, 2008, 5:35 am

sample cuda problems in finance

May 29th, 2015, 10:42 am

I would recommend this slide deck for an explanation of the space of SIMD/SIMT/XXXX possibilities and what they mean.
 
User avatar
mj
Posts: 3449
Joined: December 20th, 2001, 12:32 pm

sample cuda problems in finance

June 11th, 2015, 9:42 am

the development took place over 3 years but it was only a small fraction of what I was doing. I taught several subjects, wrote two books, did a fair amount of admin and wrote several papers on other topics at the same time. I haven't detected enough interest to get motivated about writing a book on this topic.
 
User avatar
AlexEro
Posts: 203
Joined: November 25th, 2014, 4:27 pm
Location: Ukraine
Contact:

sample cuda problems in finance

September 25th, 2015, 5:47 am

As there are important differencies between CUDA capabilities in various versions of CUDA (drivers), as well as there are incompatibles between them, I think this table of drivers' versions will be useful for every CUDA developer:CUDA 7.5 (7.5.18): driver 352.86CUDA 7.0 (7.0.28): driver 347.62driver 350.12 recommendedCUDA 6.5 (6.5.14): driver 340.21driver 341.44 recommendeddriver 344.11 recommendedCUDA 6.0 (6.0.14): driver 331.17driver 331.62 recommendeddriver 332.88 recommended (developer's version)CUDA 5.5 (5.5.20): driver 319.37driver 320.57 recommended (developer's version)CUDA 5.0 (5.0.35): driver 304.54driver 306.94 recommendedCUDA 4.2:driver 295.31CUDA 4.0:driver 270.81CUDA 3.1:driver 256.xCUDA 3.0:driver 195.xCUDA 2.3:driver 190.x
Last edited by AlexEro on September 24th, 2015, 10:00 pm, edited 1 time in total.
 
User avatar
Cuchulainn
Posts: 62395
Joined: July 16th, 2004, 7:38 am
Location: Amsterdam
Contact:

sample cuda problems in finance

September 25th, 2015, 10:58 am

For the great unwashed like myself could you explain all these numbers? On a Tesla C20270 and CUDA 5.5 we report speedups in the range [30,150] for Asian and lookback options using the MLMC method. From your chart CUDA 5.5 has quite a numbers of 'updates'/improvements?
Last edited by Cuchulainn on September 24th, 2015, 10:00 pm, edited 1 time in total.
ABOUT WILMOTT

PW by JB

Wilmott.com has been "Serving the Quantitative Finance Community" since 2001. Continued...


Twitter LinkedIn Instagram

JOBS BOARD

JOBS BOARD

Looking for a quant job, risk, algo trading,...? Browse jobs here...


GZIP: On