ok I though bx limiter was a preset. Also old plates are not very good, because they feature a short kernel for harmonics. If you need plates use vnxt ones with cuda, or start testing a clean long tailed one. When you test cuda start from something obvious, if it works tune other items accordingly
Old plate 3 has +/- 3600ms for normal and 1000ms for harmonics. That isn't short, right?
But it are two different ones, and I thought you said somewhere Cuda works best if all the kernel sizes are the same.
I was looking for a 44.1 khz preset everyone could use to test with (as in, Acustica supplied or free 3rd party demo). I don't own any VNXT stuff for example, only have their 140-plate demo, which is at 96khz.
Only reason for me wanting a 44.1khz one is that loading Nebula and cloning the tracks takes some time, and if the kernels need to be resampled it only takes longer .
Core of my 'story' was though that it actually clearly works, even if the kernel-sizes weren't 'optimal'. It's not night and day , but (as stated earlier) an improvement from 2.4x realtime to 3.4x realtime is an improvement for sure. Now if there are programs that _could_ work better because of a better kernel size, that makes it only nicer now does it .
yes start from something of the same size, better a single kernel. We need to measure cuda performances alone. It is not night and day. Cuda provides around 10x for 3d rendering, so 3x for a realtime usage is pretty good
Tried again at 96.0 khz (still 1024 ASIO buffer) with my favorite 'EAR Rooms Guitar-room'. Single kernel, 3600ms.
On Cpu alone I _thought_ I had 7 instances (single instance on 7 tracks, nothing else) working. CPU usage was showing 80% when stopped, 92% when playing. When actually listening it was stuttering like mad.
I needed to go down to 5 tracks / instances before the stuttering went away (70% CPU usage now). After playing for 10 seconds, the stuttering came back. So in the end I needed to go down to 4 tracks / instances to make and keep it smooth in playback. CPU usage was around 45% to 55% now, so room for other normal plugs .
Now, with CUDA, I ended up on 6 tracks / instances that played smoothly, but CPU usage was around 70%. When I tried to go to 7 instances, it stuttered like mad (even though my CPU was still only 75% taxed). So I think I ran out of GPU power there.
The real test would be if I could run more other / regular plugins next to a few Cuda-nebula instances or not.
My mixing template always had a EAR-guitar room aux set up ready to go and a old-plate-3 from the Acustica library. Those are now Cuda instances by default, because it seems to work OK and I can still load my MFeQ and other cpu-nebula instances .
Like Giancarlo said, it's not night and day a difference, but there IS a difference. And the less CPU power you have, the more the gains are. But you just can't use it for everything.
so I think getting a couple of Xeon's in a workstation build to go all out will still give you more power and working-pleasure in the end compared to a high-end videocard.... for Nebula stuff that is.
You know a board you can put two i7's on? Cause as far as I know, multiple-cpu-socket support is always reserved to the Xeon line, and then even the higher models .
Multiple sockets also means ram-per-socket, meaning if you want to have 16gb available to CPU, you need 32gb installed. There is a reason multi-socket is meant for the server world, it's a pain in the ass to work with and build systems for .
Still, seeing that Cuda works for Nebula in certain cases.. my mind keeps wondering what would happen with OpenCL support on cheap AMD APU's. They have very little CPU power compared to an i7, but with the help of the assisted GPU power it might be enough to run a few reverbs, which makes it handy for embedded devices.
Nebula already uses Intel IPP for a lot of stuff right? I assume because of the ipps libraries installed and you claiming it works nicely on latest I7's (AVX/AVX2 support I assume). I thought the latest Intel IPP stuff included OpenCL support. Could be a nice experiment to see what the integrated Intel graphic chips and the integrated/included (or even high-end) AMD chips can bring to the table.
An an off-topic side-note: Did you guys ever test what impact triple channel and quad-channel memory had for Nebula? (In other words, might there be a reason to go to the LGA2011 range instead of the LGA1150 range).
When Sandybrdige-E came out everyone was comparing if it would actually bring extra speed to the table compared to the Sandybridge and Ivybridge i7's at the time. In every single benchmark, there was NO performance gain... except for Photoshop. There the extra memory bandwidth did help apparently . Always wondered if Nebula had anything to gain with more memory bandwidth.
I have no luck with cuda. Installed new cuda toolkit 6.037 last cuda 64 breadge necubr03.dllsits on SYSWOW64 Cuda do not work. can not switch to 15PPS - have a ! mark. the card i have is Nvidia quadro k4000. Please help. What is wrong? Do I need to install cuda bridge file in System 32 too?