Tag Archives: CPU

GREAT Tips on Python CUDA GPU

GREAT Tips on Python CUDA CPU

This came in from the same Meetup member who suggested this computer

How to build a trading computer

Thanks to this person

 

This one has 16 x 512 cores ( or 8192 cores)
at 1.3 GHz. (Normal pc has like 8 or 12 cores)

If the code is cuda enabled that's more cores than than the caltec hpc lab.
 (Python has libraries that plugin that are coded to use cuda)

US DOE department of energy is said to be switching to cuda for their next super computer 
for energy modelling. (100-300 petaflops)

Right now the only cuda hpc lab available is Nvidia. (They do have a free trial you can
 subscribe to) To make it comparable, you'd have to rent 1500-2000 quad core machines to be able to make a comparison.

(Check out Nvidia k80 and Nvidia jetson)
K80 is what physicist use to model nuclear reactions, and oil engineers for fracking 
calculations) (jetson is what's being used for visual processing for automated driving)

Keep in mind this is only for research and development aka model building not actual
 trading. (But it will definitely speed things up)

Join my FREE newsletter to learn more about these tech tip to enhance your automated trading

NOTE I now post my TRADING ALERTS into my personal FACEBOOK ACCOUNT and TWITTER. Don't worry as I don't post stupid cat videos or what I eat!

What is your opinion of this Cheap Ubuntu Linux CPU cluster for your quant trading model generated by Matlab or C++ ? Affordable FPGA?

What is your opinion of this Cheap Ubuntu Linux CPU cluster for your quant trading model generated by Matlab or C++ ? Affordable FPGA?

Discuss here:

http://quantlabs.net/academy/forum/technical-forum/cheap-ubuntu-linux-cp-cluster-for-your-trading-model-generated-by-matlab-or-c-affordable-fpga-this-product-came-from-london-based-quant-httpshop-adapteva-comcollectionsp/#p59

Join my FREE newsletter to see if I might use this cluster

NOTE I now post my TRADING ALERTS into my personal FACEBOOK ACCOUNT and TWITTER. Don't worry as I don't post stupid cat videos or what I eat!

what do you’ll think of a CPU, GPGPU, FPGA hybrid computing platform..any body working on something like this?

what do you’ll think of a CPU, GPGPU, FPGA hybrid computing platform..any body working on something like this?
==Have worked with custom IC design before
==Mixed processor technology (uP/FPGA/GPU) systems have the potential for delivering extraordinary performance givine their modest SWaP (size, weight, and power) footprint. The econonmics of these systems is also compelling– multi-TFLOP’s in a $10-$15,000 package is hard to ignore.
But this combination of high performance and small SWaP comes at a price; they are notoriously diffiuclt to program requiring multiple technical specialists on the application development team. Still, new products such as our hprcARCHITECT make using them much easier than with traditional developmnet approaches.
I put together a system for delivery to the US Air Force that combined a motherboard with two Intel i7 uP’s, four Xilinx FPGA’s, and two nVIDIA TESLA GPU’s. To say the very least– it screamed. It also demonstrated that there is no performance difference between Windows and Linux for high-performance applications. But Windows code is significantly easier/faster to get into production giving it an edge in total time to first solution.
==I worked on FPGA board + PC implementation before. One can easily speed up the overall system performance by moving some lower-level processing into FPGA chip. The capability of embedding a CPU inside of the FPGA and having some control program running in the FPGA also provide some further flexibility. However, debugging in this kind of system is very troublesome and it is much more likely to have bugs in a hybrid system than the regular software environment.
==I agree with James– hybrid systems will, initially, be buggy. But this can we dealt-with by choosing the right development tools and working from a good specification to begin with. Hybrid systems are no place for “hackers”– they take solid engineering and well designed tools.
More development projects are buggy due to a poorly engineered software specification (including theĀ ==
Funny that – we filed a patent on this in December of last year…

NOTE I now post my TRADING ALERTS into my personal FACEBOOK ACCOUNT and TWITTER. Don't worry as I don't post stupid cat videos or what I eat!

Talk on how C++, Matlab, R to accelerate a CPUI using FPGA/CUDA

Talk on how C++, Matlab, R to accelerate a CPUI using FPGA/CUDA

Some important features of fermi that require highlighting:

– True Floating Point – 754-2008 standard, 16-bit floating point memory format, multiply-add support
– ECC – hard requirement these days to stave off SEU, one can’t have a bit flipping when dealing with a million dollar transaction
– Switching between 32 and 64 bit addresses is streamlined
– Scratchpad, GPU and system memory all reside on same 64-bit interface -> this makes compiling C++ a breeze
– Atomic instructions avoids any other process from overwriting memory values, great handshaking
– Allows kernals to overlap executions

I do agree with Alex on the lack of C++ constructs being supported, its just a matter of time before that changes (this is all relatively new in the past 5 or so years, it took FPGAs much longer to finally streamline their process), so those with vision and who can see where this is all heading, and has already headed for that matter, its best to be among the first on the train then left at the station. So if you have the ‘time’ and ‘$'( I will argue after all the required libraries you will need with Matlab, if you choose that route, you could have armed yourself with a serious fermi device), its in your best interest to learn how to program some of your exercises onto a GPU, again if you have time. That will differentiate yourself when applying to jobs, everyone can use Matlab, R, code in C++ etc, not everyone can use the CUDA architecture to offload // processing…

For those in doubt, download folding at home and see for yourself the advantages you get from a NV compute device versus porting it to a CPU(download both CPU and GPU version for comparison purposes):

Front Page

Heck, code up some Monte Carlo simulations and set your number of sims into the millions or better yer price out every option in the market via black-scholes or using any binomial/trinomial method and you will quickly see what performance advantage due to mass parallelism you gain with GPU vs CPU. Its almost an unfair contest.

R is your best choice for the programs you outlined. Its free, open-source and quite powerful.

This was found to be part of Linked In Group Conversation.

NOTE I now post my TRADING ALERTS into my personal FACEBOOK ACCOUNT and TWITTER. Don't worry as I don't post stupid cat videos or what I eat!