Cytomic 'The Glue' Compressor

[Tp3]

ETA on Glue HD?

Maybe call it "Adhesive" ?

Or call it 'The Superglue'

Hey, this is naissse

So, let him release a saturation plugin, and bundle it with SuperGlue and call the collective "Adhesive"

[MogwaiBoy]

The Adhesive Bundle

You genius! Our work here is done, Tp3. Good work team haha

[Tp3]

Now we wait

[MogwaiBoy]

Our next project will be the long awaited follow-up to our award winning Emotional Kazzoo sample library.

[andy-cytomic]

Stam Audio (insanely great products if you can live with missed deadlines) has an SSL clone coming out with the option to add optional API-style 2520 op-amps or Neve style transformers to the signal path. Not sure if that kind of thing is of any interest to you, but it might be another nice value-add to do something similar.

http://www.stamaudio.com/sa-4000.html

Good idea! I've emailed them to ask about kit versions that I could measure each component before soldering it on. That would save a lot of guesswork

[softska]

Andy, is support on holiday? My Glue stopped working on my system again, but multiple email to support just radio silence...

[Aloysius]

Hello, Andy? I'm calling from the 1980s.

[sleepcircle]

Is andy ok?

it might be stupid of me to suddenly feel worried, but he's been weirdly silent for about two months. That last post a couple steps above mine was the last time he posted here, he's silent on twitter, and if he's not replying via email, either…?

idk. like, disclaimer: i'm an extremely nervous person. i frequently imagine that people "have died" if they're a couple hours late coming home, or something. I guess I'm just asking if anyone else knows something or has heard from him.

[zvenx]

Is andy ok?

it might be stupid of me to suddenly feel worried, but he's been weirdly silent for about two months. That last post a couple steps above mine was the last time he posted here, he's silent on twitter, and if he's not replying via email, either…?

idk. like, disclaimer: i'm an extremely nervous person. i frequently imagine that people "have died" if they're a couple hours late coming home, or something. I guess I'm just asking if anyone else knows something or has heard from him.

He has posted since then on gearslutz..... he usually goes quiet when he is really busy working.

This was 3 weeks ago.
https://www.gearslutz.com/board/showpos ... tcount=194

rsp

[sleepcircle]

That's a small relief. Thank you.

[andy-cytomic]

Busy working away here and making very solid progress. I had around a month of downtime with the family getting a stomach bug, then me getting it, then all of us getting colds, then a week's holiday, but back into things now.

Warning: maths talk below!
--
I'm in the final stages of a big re-write of the entire circuit solver, using a new symbolic expression class and sparse matrix solver with cross matrix optimisations so I can handle tacking larger and more complicated circuits (ie bigger and faster solving of equations with lower cpu results for final simulations).

I few months ago I invented the theory behind a new parallel low cpu solving method I'm calling the PDK method (Parallel Discrete Kircfhoff's method - was going to call it SDK but that already stands for software development kit!). This method can make use of the wide vector operators of modern CPUs to solve highly non-linear circuits efficiently and in parallel. Yesterday I had a huge win, I implemented my new PDK method to solve the main drive section of The Scream efficiently and accurately. Normal circuit solving requires up to 16 iterations per sample, each of which has around 12 calls to exp (x) (the exponential function), and the number of iterations is higher for signals with high frequencies like noise - so it is not constant (unless you want to always chew cpu and do 16 iterations all the time!). The new method only needs 3 iterations, and only has 4 calls to exp (x), so it's around 16 times more efficient for worst case signals. This is the difference between taking 16% of your cpu vs 1%.

The future is also very promising because this method should scale well for use on GPUs - as long as the main memory to GPU memory speed is fast enough to handle the audio streaming to and from the CPU, since the inter kernel data communication is quite small. This is a "big deal" since most audio people have GPUs sitting there doing nothing, imagine if I could fully utilise them so every channel on your DAW could have multiple detailed analog models and your CPU just had to be a fancy disk and audio interface controller!

But first to get PDK up and running using AVX 256 in The Scream, and then in The Glue HD mode

[andy-cytomic]

Andy, is support on holiday? My Glue stopped working on my system again, but multiple email to support just radio silence...

I don't always check forums for support requests, but we do handle support as quickly as possible via email. Can you please post here to let people know when your support request was handled, or possibly let me know your Cytomic username so I can look it up and let people know?

Sometimes there is a delay if you email support on Friday since we are GMT+8, so Friday day time in USA is already close of business in Perth Australia, which means you possibly won't get help until Monday.

[Funkybot's Evil Twin]

Busy working away here and making very solid progress. I had around a month of downtime with the family getting a stomach bug, then me getting it, then all of us getting colds, then a week's holiday, but back into things now.

Warning: maths talk below!
--
I'm in the final stages of a big re-write of the entire circuit solver, using a new symbolic expression class and sparse matrix solver with cross matrix optimisations so I can handle tacking larger and more complicated circuits (ie bigger and faster solving of equations with lower cpu results for final simulations).

I few months ago I invented the theory behind a new parallel low cpu solving method I'm calling the PDK method (Parallel Discrete Kircfhoff's method - was going to call it SDK but that already stands for software development kit!). This method can make use of the wide vector operators of modern CPUs to solve highly non-linear circuits efficiently and in parallel. Yesterday I had a huge win, I implemented my new PDK method to solve the main drive section of The Scream efficiently and accurately. Normal circuit solving requires up to 16 iterations per sample, each of which has around 12 calls to exp (x) (the exponential function), and the number of iterations is higher for signals with high frequencies like noise - so it is not constant (unless you want to always chew cpu and do 16 iterations all the time!). The new method only needs 3 iterations, and only has 4 calls to exp (x), so it's around 16 times more efficient for worst case signals. This is the difference between taking 16% of your cpu vs 1%.

The future is also very promising because this method should scale well for use on GPUs - as long as the main memory to GPU memory speed is fast enough to handle the audio streaming to and from the CPU, since the inter kernel data communication is quite small. This is a "big deal" since most audio people have GPUs sitting there doing nothing, imagine if I could fully utilise them so every channel on your DAW could have multiple detailed analog models and your CPU just had to be a fancy disk and audio interface controller!

But first to get PDK up and running using AVX 256 in The Scream, and then in The Glue HD mode

So 16% CPU becomes 1% CPU? That sounds huge! Please begin emulating everything at a level of quality and efficiency that other developers won't be able to compete with. My credit card will be ready to go.

How would the GPU be with latency? Would it add a considerable bit like the UAD cards? Or, are you saying that the latency isn't as bad because the GPU's are already running lean because they're so tightly integrated into the kernal of the OS?

[sleepcircle]

sorry about being the forum userbase equivalent of a clingy girlfriend.

also that other stuff you posted about how the programming is going is fascinating, i've only ever read about this sort of stuff being used in rendering engines and simulations and stuff.

although i mean, this technically IS a simulation. i guess it just doesn't operate in the same domains as most other simulations I can think of.

[andy-cytomic]

So 16% CPU becomes 1% CPU? That sounds huge! Please begin emulating everything at a level of quality and efficiency that other developers won't be able to compete with. My credit card will be ready to go.

How would the GPU be with latency? Would it add a considerable bit like the UAD cards? Or, are you saying that the latency isn't as bad because the GPU's are already running lean because they're so tightly integrated into the kernal of the OS?

Yes, x16 faster in the worst case (edit: for the regular solver) for a high frequency noisy signal where the previous state isn't useful to guess the next state. For a more real world noisy signal like white noise then it should be around x10 quicker than brute force solving like I'm currently doing in The Scream. All these numbers are guesses, I need to profile things to see the actual speedup, so I should have some more accurate numbers in around a month.

The UAD cards have the same sort of issues as GPU processing will in terms of latency, but typically GPUs get the fastest PCIe slot on the system, so they could be out in front, but I really don't know at this stage. The latency will depend on the memory throughput of system CPU RAM to GPU RAM, each audio buffer will have to be sent to the GPU memory for processing and then sent back to the main system memory so the DAW can use it. This could even make underpowered integrated graphics cards possibly faster than external ones as they use system memory directly. Audio is so very different to graphics, so I really have no idea what to expect with regards results!

I will need to write a test application that just pumps random numbers to the GPU and then returns them all again in another buffer straight away to really test what is possible in terms of latency. The GPU stuff is a long term thing, I will release multiple plugins using AVX 256 first, which can do 8 x float32 in a single SIMD operation. For me this equates to 4 lumped stereo non-linearities all computed at once.