SERVING THE QUANTITATIVE FINANCE COMMUNITY

  • 1
  • 6
  • 7
  • 8
  • 9
  • 10
 
User avatar
Cuchulainn
Posts: 62950
Joined: July 16th, 2004, 7:38 am
Location: Amsterdam
Contact:

sample cuda problems in finance

October 15th, 2015, 4:49 am

QuoteOriginally posted by: AlexEroQuoteOriginally posted by: CuchulainnQuoteIsn't the problem of parallelizing these PDEs tied to the topology of the mesh and the need for each core access values in mesh-adjacent cores? I get the sense that the core-to-core accessibility and bandwidth inside a GPU is particularly poor. Yes. In particular, we can get false sharing when 2 threads modify data stored on the same cache lines.PCI-E bus transfers from/to host<->GPU can kill any algorithmic speedup. The overhead is always 1-5 ms per cudaMemcpy, and cannot be reduced. Hello, Intel.Are you saying that GPUs are insensitive to Amdahl's law? If you have a high serial fraction then no h/w will speed up the algorithm, which is inherently sequential.
Step over the gap, not into it. Watch the space between platform and train.
http://www.datasimfinancial.com
http://www.datasim.nl
 
User avatar
AlexEro
Posts: 203
Joined: November 25th, 2014, 4:27 pm
Location: Ukraine
Contact:

sample cuda problems in finance

October 15th, 2015, 10:20 am

QuoteOriginally posted by: CuchulainnQuoteOriginally posted by: AlexEroQuoteOriginally posted by: CuchulainnAre you saying that GPUs are insensitive to Amdahl's law? If you have a high serial fraction then no h/w will speed up the algorithm, which is inherently sequential.Nope. I did not say *that*. I am saying that almost any old algorithm can be re-designed to fit GPU with proper speedup. Even more: any MATH problem can be solved by various methods. Who cares that you are using "an old, outdated, less precise, forgotten" method "from 1970"? - if it gives (surprise!) 100X speedup on GPU? Time to dig into old math books, re-considering of all old algos.
 
User avatar
Polter
Posts: 2526
Joined: April 29th, 2008, 4:55 pm

sample cuda problems in finance

October 15th, 2015, 10:44 am

QuotePCI-E bus transfers from/to host<->GPU can kill any algorithmic speedup. Right, hence the "accelerator attach" trend (the move to On-Package to On-Chip to On-core).May have to get (re)acquainted with FPGAs, though:http://www.ece.cmu.edu/~calcm/carl/lib/ ... upta.pdfAt the same time, there are also productivity-oriented (in a relative sense) solutions familiar from the GPGPU world:http://www.xilinx.com/about/xcell-publi ... erestingly, as the above illustrates, it's not limited to just Intel-Altera; Xilinx-Qualcomm ARM partnership is worth observing, too:http://semiaccurate.com/2015/10/14/qual ... x-servers/
Last edited by Polter on October 14th, 2015, 10:00 pm, edited 1 time in total.
 
User avatar
Cuchulainn
Posts: 62950
Joined: July 16th, 2004, 7:38 am
Location: Amsterdam
Contact:

sample cuda problems in finance

October 15th, 2015, 11:18 am

QuoteOriginally posted by: AlexEroQuoteOriginally posted by: CuchulainnQuoteOriginally posted by: AlexEroQuoteOriginally posted by: CuchulainnAre you saying that GPUs are insensitive to Amdahl's law? If you have a high serial fraction then no h/w will speed up the algorithm, which is inherently sequential.Nope. I did not say *that*. I am saying that almost any old algorithm can be re-designed to fit GPU with proper speedup. Even more: any MATH problem can be solved by various methods. Who cares that you are using "an old, outdated, less precise, forgotten" method "from 1970"? - if it gives (surprise!) 100X speedup on GPU? Time to dig into old math books, re-considering of all old algos.Too general to respond to.I can't remember any good algos from the 70's. The golden age was the 60s.
Step over the gap, not into it. Watch the space between platform and train.
http://www.datasimfinancial.com
http://www.datasim.nl
 
User avatar
AlexEro
Posts: 203
Joined: November 25th, 2014, 4:27 pm
Location: Ukraine
Contact:

sample cuda problems in finance

October 15th, 2015, 11:29 am

QuoteOriginally posted by: CuchulainnQuoteOriginally posted by: AlexEroI can't remember any good algos from the 70's. The golden age was the 60s.I agree with you.Besides, I am not a nephew of Putin. Why are you so agressive with my posts here? Or, maybe you are so agressive all the time?
Last edited by AlexEro on October 14th, 2015, 10:00 pm, edited 1 time in total.
 
User avatar
Cuchulainn
Posts: 62950
Joined: July 16th, 2004, 7:38 am
Location: Amsterdam
Contact:

sample cuda problems in finance

October 15th, 2015, 11:53 am

QuoteOriginally posted by: AlexEroQuoteOriginally posted by: CuchulainnQuoteOriginally posted by: AlexEroI can't remember any good algos from the 70's. The golden age was the 60s.I agree with you.Besides, I am not a nephew of Putin. Why are you so agressive with my posts here? Or, maybe you are so agressive all the time?Vladimir is a very good judoka. I think more world leaders should do martial arts. BTW what's the best 70s' algo? or is that a too aggressive question?
Last edited by Cuchulainn on October 14th, 2015, 10:00 pm, edited 1 time in total.
Step over the gap, not into it. Watch the space between platform and train.
http://www.datasimfinancial.com
http://www.datasim.nl
 
User avatar
AlexEro
Posts: 203
Joined: November 25th, 2014, 4:27 pm
Location: Ukraine
Contact:

sample cuda problems in finance

October 15th, 2015, 1:05 pm

QuoteOriginally posted by: CuchulainnQuoteOriginally posted by: AlexEroQuoteOriginally posted by: CuchulainnQuoteOriginally posted by: AlexEroI can't remember any good algos from the 70's. The golden age was the 60s.I agree with you.Besides, I am not a nephew of Putin. Why are you so agressive with my posts here? Or, maybe you are so agressive all the time?BTW what's the best 70s' algo? or is that a too aggressive question?Yes. Yes!. THIS IS a very agressive question! In order to show my "ignorance" You pretend that You dont know that modern programming (who needs algorithms without programming?)started from UNIX and C programming. UNIX was the best algorithm of modern times! The C language compiler was the best algo of modern times! Why ? Because without those "algos" there will be no all the others.And you pretend that you dont know that ***ing (self-censorship) OS' system time (still) started here and there ...... from 01-Jan-1970. I am very far from the idea that you really dont know that.
Last edited by AlexEro on October 14th, 2015, 10:00 pm, edited 1 time in total.
 
User avatar
Cuchulainn
Posts: 62950
Joined: July 16th, 2004, 7:38 am
Location: Amsterdam
Contact:

sample cuda problems in finance

October 15th, 2015, 3:49 pm

QuoteUNIX was the best algorithm of modern times! Last time I looked was UNIX an OS, but hey, what the heck!let's call it a draw.
Step over the gap, not into it. Watch the space between platform and train.
http://www.datasimfinancial.com
http://www.datasim.nl
 
User avatar
Polter
Posts: 2526
Joined: April 29th, 2008, 4:55 pm

sample cuda problems in finance

October 15th, 2015, 5:10 pm

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