One of the developers contributing to NAM (Mike Oliphant) made an LV2 version in order to have a version of NAM working for LInux, less features compared to the original but the important part - loading nam models - is fully working.
MOD assisted in getting this LV2 plugin version working as expected by most hosts, and after some testing we are pushing the plugin to the store!
The plugin requires the latest v1.13.2 for being able to upload NAM files on the correct directory.
Note that old regular “heavy” models do not work, models must be of nano type.
Tonehunt uses the #nano-mdl tag for this purpose #nano-mdl Models | ToneHunt
We uploaded a few models of this type already to it.
Enjoy!
PS: more content related to this plugin coming very soon
At first glance it sounds cool. (I literally mean the sound itself). I am starting to suspect that the next generations of the FX processors should have tensor cores on board to compete.
The latest rockchip based ARM boards already have that in mind, with optional NPU cores.
I do not know anyone who has experience working with such processors, so got no one to ask about details on how they work exactly. But it seems a worthwhile investment, the stuff these cores allow to do match what latest AI modeling does (basically dedicated cores to handle convolution network operations, in between other things).
We looked into it a bit inside MOD, but only quickly to understand that indeed might be useful.
At least on rockchip side the stuff we care about does not have a C/C++ side ready, only python, so it wouldn’t be useful for realtime processing even if we had one in our hands right now.
It is something to keep an eye out for, there is potential there but will take some time.
Who knows if at some point they become so commonplace it ends up being part of the ARM spec, alike the FPU now being mandatory on all ARMv8+ and arm64 hardware
Good news for DuoX users - standard nam files works great, the CPU load with IR CAB loader is below 60%. I’ve loaded free pack from tonejunkie - heaven!! They sounds fantastic. Bad news for Dwarf users, as mentioned earlier, they are too heavy.
The good thing about NAM is that there are few thousands of free profiles plus there are at least 4 places when you can buy profesional captures. This is what is missing with Aida but AIda is quite new profiler and the base of prfiles will rise. What is more Aida is about to release 3 AI models based on egnater, fender and Marshall with knobs working as in real amps, so this is something NAM and other prfilers do not have.
I’ve checked NAM with dwarf yesterday - MOD NAM profiles works but it is not possible to use IR CAB loader since CPU can not handle it. It works with MOD cabs like Vintage.
Usually the NAM nano profiles take 67% of CPU of the Dwarf, which leaves enough space for the IR loader + Reverb. Which is minimal but somehow a comlete rig for guitar.
U didn’t manage to make that work?
I wonder if in the case of ground loop compensation on, meaning with the CPU downclocked to 1.2GHz, it leads to not being able to have nam + cab. It might be just on the limit so maybe @Lukasz is using that approach?
Wow!!! it’s a CPU killer!!! on Dwarf, a pedalboard with Mod Auto Input, Neural Amp Modeler (loaded with nano-mdl), Mod IR Loader Cabsim (Loaded with 30K of waw IR) goes to 99.4% of CPU… unusable…
Without Cabsim it goes to 72.7%… unusable!!!
it is by design, NAM developers prefer to have this quality of sound without any kind of compromise.
LSTM based models can run at around 35% on a Dwarf, so basically half of the current WaveNet nano approach, if sound ends up “degraded” is a bit questionable.
The point of entry is getting the plugin to run, which is the case now.
Optimizations can still be made, but yeah with these kind of plugins the pedalboard is always going to appear a bit empty compared to other more complex setups.
It is a case of being smarter vs brute-forcing the problem. AI tech typically goes for the brute-forcing approach, throwing huge amounts of CPU at it first, then we optimize things by comparison until a good compromise is achieved.
I think we need to do an LSTM vs WaveNet blind test, as the NAM plugin supports both so it won’t be a case of different engines (like a possible AIDA-X/RTNeural vs NAM comparison)
I’m thinking about to try to profiling the entire section od/dist/amp/speaker to see if the result is good, maybe it can help to have enough CPU to use some dly/rev after that profile.
I saw that with AIDA-X the better way was to profile only the amp but I never tried to profile all this section on NAM, do you have any suggestions, hints?
Hi, I have a Duo and a Dwarf… can someone please explain how to load NAM models into the NAM virtual pedal? My Duo is version 1.13.1.3312, how do I get 1.13.2, or is that only for the Dwarf? And I downloaded a free NAM pack from Tonejunkie, so how do I get it into the NAM? Sorry, this is completely new to me…
EDIT: Okay, I think I figured it out. First of all, I have to manually update the Duo to version 1.13.2 as automatic updates for Duo are no longer supported. And then I should be able to load the NAM file using the File Manager to the correct folder (which requires updating to version 1.13.2).
I’m having trouble updating, but hopefully it works shortly!
Wavenet models will not work on the old Duo, not even nano models, poor thing cannot keep up.
it can still load LSTM-16 models, but as far as I know there are none of that type made for NAM right now.
The Dwarf can load wavenet nano models, with roughly 66% cpu usage so it doesn’t leave a lot of space for much else at the moment.
We are discussing internally possible optimizations and paths to take in order to get other plugins in the chain at the same time.
A mono cabinet IR loader that can do multi-threaded processing seems to be the way to go, you can already do some experiments for this by uploading cab IRs in the reverb folder and loading them with the x42 convolution plugin. this plugin is not really suitable for cabs though, dry/wet for one doesn’t make sense on cabs…
We will have news on this soon.