Indeed Knights Corner is not shipping yet. But, the previous generation Knights Ferry prototype has been shipped in SDVs to a number of teams around the world.

<t>QuoteOriginally posted by: CuchulainnQuoteOriginally posted by: twofishQuoteOriginally posted by: CuchulainnSo, the question is: which applications must be in double precision and which are OK in single precision. I doubt one size fits all. Till now, we used double but a nice try would be to do a...

<t>QuoteOriginally posted by: Traden4AlphaTry incrementing a float by 1 in a loop that runs 10^8 iterations.Right; these are the things we always have to think about. Eg. this output should be deeply unsurprising on an x86.32b 10.000001907348632812500000000000000000000000000000000000000064b: 9.99999...

<r>QuoteOriginally posted by: CuchulainnQuoteYou should already, for CPUs, be thinking of algorithms that use lower precision where possible and refine with higher precision as necessaryI think this is a non-starter. Algos are very difficult at the best of times without bracketing the code to make s...

<t>QuoteOriginally posted by: CuchulainnQuoteSequential equivalence - sure. You won't get the same answer - you need to make sure its an equally useful answer. Same goes for alternative sequential implementations. Just naively doing everything in the "obvious" order on a sequential machine isn't as ...

<r>QuoteOriginally posted by: CuchulainnQuoteOpenCL supports task parallelism via the enqueue task api. In 1.0 the host builds a task graph, using event dependencies, and submits it for execution. That can be on CPU or GPU or both. First time around, there are some restrictions like tasks not being ...

<t>QuoteOriginally posted by: CuchulainnI took an adapted example from the undergraduate textbook by Fox and Mayers (Computing Methods for Scientists and Engineers, 1968) to see if could force round off errors. There is a difference.. By modifying 'fac' get different values for floats and double (an...

<t>QuoteOriginally posted by: CuchulainnQuoteOriginally posted by: zetaJawabean, to your specific questions;a) Not all algos are made for GPU; it is SIMD or data-level parallelizationa) function lengths are strongly limited by host-device exchanges, minimize the latter ie., reuse data at the device ...

<r>QuoteOriginally posted by: CuchulainnQuoteBut I've definitely found best performance from writing cuda directly. I have less experience with opencl - it has a lag behind the features of cuda.Ideally, OpenCL would be a more portable interface in the medium to long term but I suppose native interfa...

<r>For best performance, you're going to have to learn what's happening. There are tools on the way to help from Nvidia, such as Nexus:<URL url="http://developer.nvidia.com/object/nexus.htmlYou">http://developer.nvidia.com/object/nexus.htmlYou</URL> can try to take a low impact approach if you have ...

<r>QuoteOriginally posted by: CuchulainnQuoteI don't think you can just refactor non-GPU code in a straightforward way and make it run as fast as possible in a GPU. These GPUs have their own optimizing rules. To get those spectacular performance increases you must be very careful about optimizing th...

<t>QuoteOriginally posted by: CuchulainnI am afraid I cannot answer that question directly but a similar test might be a calculation that takes 4 seconds reponse time how would this fare on a cluster?Too many unknowns I'm afraid..(even before thinking of going to a cluster I'm trying to gauge if the...

<r>QuoteOriginally posted by: CuchulainnQuoteOriginally posted by: eredhuinI get that OpenMP gives you finer control over what the compiler actually does, but (aside from the clustering product), do I care, if the compiler can find all the candidate loops for me?I reckon that the compiler builders c...

<t>QuoteOriginally posted by: eredhuinThe good news is that it's very easy to use MPI. Yes, its easy in the sense that chainsaws and C are easy. If you have no problem with that then yes you'll wonder why the fuss about using it. Given the goat$%&^s that people can make of even minuscule (< 50 k...

<r>QuoteOriginally posted by: eredhuinOpenMP is a shared-memory construct only so you may want to mix and match MPI for clustering.That is indeed the by now a reasonably common approach to dealing with a cluster of SMPs. Intel have extended OpenMP to distributed memory though, with their Cluster Ope...