CUDA knowledge - does it really increase employee's value?

finanzmaster · Post by **finanzmaster** » Tue Feb 01, 2011 10:34 am

Hello together,

There was pretty much talks (hype?) on CUDA, there are some topics at this forum as well.

But how widely does one engage CUDA in financial world?

Classical example is BNP Paribas.

But as to quant jobs in Germany I, for one, have seen [so far] only one job, where CUDA knowledge was desired (it was not even a must-have).

Probably, there is more demand for CUDA in City or NY?

FDAXHunter · Post by **FDAXHunter** » Tue Feb 01, 2011 10:43 am

Dude.. CUDA is just one small tool in a very large toolbox. Not knowing CUDA will not hurt you. Just as knowing one more language will not really help you.

finanzmaster · Post by **finanzmaster** » Tue Feb 01, 2011 11:39 am

I would not oversimplify so much.

CUDA is not yet another programming language, it is a paradigm (of programming parallel GPUs).

I mean the following: when one comes from C to C++, one learns not a new language but the new concepts: OOP, template programming - and it is much more.

Moreover, CUDA can increase computational performance upto 100 times, so it is not "just one small tool", it can be a breakthrough.

But does anyone really need it? In high-frequency trading very likely.

On the other hand, if a bank must satisfy a regulator and the latter says 1000 Scenario-simulations will do, why should a bank bother trying to perform 100000 simulations in the same time?

So I am curious how many employers do really care. As I said, in Germany there are not many. So what about the City or NY?

FDAXHunter · Post by **FDAXHunter** » Tue Feb 01, 2011 12:06 pm

I'm telling you, it is one more tool.

pj · Post by pj » Tue Feb 01, 2011 12:12 pm

it is not a tool it's a paradigm Cool

FDAXHunter · Post by **FDAXHunter** » Tue Feb 01, 2011 12:20 pm

Hammertime <--- Tool, you see?

pj · Post by pj » Tue Feb 01, 2011 12:51 pm

Piss <---paradigm.

did you get it?

prophet · Post by **prophet** » Tue Feb 01, 2011 6:45 pm

CUDA can increase computational performance upto 100 times

For highly vectorizable algorithms with fairly regular memory access patterns, yes. For most algorithms it depends on many factors. Many algorithms can not be easily mapped to a GPU. One must also consider the implementation cost of CUDA versus out-of-the-box GPU accelerator tools like www.accelereyes.com

CUDA is just basic extensions upon C. A good coder can master it in a few days. The real skill of interest is understanding hardware (GPU vs CPU), understanding the problem and choosing the right tools (CUDA, OpenCL, C/C++, etc) to map the problem to computational hardware at appropriate cost.

gary_badridge · Post by **gary_badridge** » Fri Feb 04, 2011 2:37 am

It's better to have a very solid grasp of the high performance languages (c/c++) and be familiar with how the problems can be parallelized or broken down (ie fluids vs for example search). High-end hardware changes very quickly, but the basics stay the same.

spaghetti_o_o · Post by **spaghetti_o_o** » Fri Mar 04, 2011 3:15 am

You can run a search on dice.com or bloomberg for "cuda finance" and see what gives, then run the same with "c++ linux finance". Or ask your favorite recruiter.

In the end it's always harder to design, implement and *debug* a robust parallel solution then to learn a specific technology or api. Also keep in mind that today you can get a 12 cores server for ~$3,000 (24 cores if hyper-threaded) and 1Gb interconnects are cheap, so it's not like we are missing parallel platform solutions.