FPGA: Cloud Compile Service compiles take very long

[Sam_Sharp]

And this is why: http://digital.ni.com/public.nsf/allkb/4501EB08F4819BFD862576120072A606

[dan_u]

Thanks, but then I understand even less why the results between local and cloud compile differ that much.

[Sam_Sharp]

---

@dan_u wrote:

Thanks, but then I understand even less why the results between local and cloud compile differ that much.

---

Probably because of the way FPGA compilations happen - it's an iterative process to find the 'best' solution according to the compilation criteria with an infinite number of possible solutions. It might be that the compilation worker on the compile farm is different to your local version or that it uses different 'starter' placements (which would produce a slightly more/less efficient design on different machines).

There's a whitepaper here: http://www.ni.com/white-paper/11573/en/

[Intaris]

If I may chime in one one point (I^m also directly affected by this issue).

The variation from compilation to compilation over all of our PCs (We have 3 FPGA licenses) tend to be in the region of single digit LUT counts (<10).  A difference of 3000 LUTs between local and cloud compile is absolutely miles away from this variation.

Shane.

[Intaris]

Another data point.

Local compiles typically take a few minutes on the "generating cores" step.  A cloud compile can take over 1.5 hours.

Is there no way the files stored locally can be transferred along with the intermediate files in order to significantly reduce this time?

The differences in resource utilisation remains.  Funnily enough the estimated resource requirement ont he cloud compiles is HIGHER whereas the actual usage is LOWER (significantly).  Is the cloud compile server using different Xilinx core components?

[JoeWork]

What version of tools does the cloud compiler use when it compiles your version?  Would it be different?  Or is there something in your design files that specifies a version to use?

We the free cloud service was introduced I thought it was a good idea until I realized that it wasn't any faster than the PC that I am running.  And I am running one that is about 1 GHz lower than yours.

From what I could determine when I ran it was that the core generation was taking the longer time.  I guess that part of it is a bottle neck in there compilation.  Is there a license limitation?  Or maybe a limitation of how many tool instances can be run at a time?

When I tried to work with tech support I got the run around about trying it a different time and on multiple computers.  It ended up not being worth the time fixing it.  You would think that NI would be more active at figuring out this issue.

I think the only reason the cloud service is useful give this limitation is for running several jobs at a time.