Neural Amp Modeler - NAM

I just finally got around to downloading/installing NAM. It really is sounding great!!! Still going through captures, but the aforementioned AC-10 is fantastic indeed! One thing I’m noticing though is the CPU. Seems to be a CPU hog. Anyone else noticed that? When I switch to racks with TH-U or BIASFX for example, my CPU drops to 2-3%. NAM racks jump to 30%.

The cpu consumption depends on the setting used to create the model: standard, lite or fether.
“lite” models will run approximately 1.5 times faster than “standard”.
“feather” models will run more than 2 times faster than “standard”.

More details on the training process (from the KVR forum):

“The more epochs a profile is trained with, can give higher quality/accuracy to a certain point, which can be determined with the ESR value/graph generated during and after training. The lower the ESR value, the more accurate it is to the original signal. There reaches a point (i.e. a point of diminishing returns) in training where the ESR wont get any lower with a higher epoch count. Usually the more harmonic content an amp has, i.e a higher gain amp or pedal/amp combination, the more it can benefit from a higher number of epochs used, to be more accurate.
There have been options added into the current trainer software version to give “feather” versions that train with less epochs for lighter/lower cpu usage, and a number of users making profiles are offering captures with both feather light versions (slightly lower accuracy/quality) and also higher epoch count for higher accuracy/quality.”

The NAM can even be used on a Raspberry Pi 4 with a “feather” model !


Thanks for the great explanation! So far I couldn’t tell the difference which settings might have been used to create the captures, as all of the 15 or so captures I downloaded ran at about the same 30% CPU consumption.

I can appreciate that the higher accuracy cycles will eat up more CPUs, but I can’t help not to think some of this is “birthing pains” due to the infancy of the software. Meaning, the big $$$ brands also use profiles and captures, but their CPU usage is 10% or less of NAM, which is totally understandable, given the resources they have available. This is still super exciting and can’t wait to see where it goes as the open source allows virtually unlimited potential for growth and improvement. The raspberry pie video is a good indication of that!

Incidentally, I just read on the FB page that the next update is supposed to cut CPU usage in half. :slight_smile:


It would be a good idea to add in the tags of the proposed models what type of capture has been used to recognize the models more or less heavy in cpu consumption.

1 Like

New plugin release! Version 0.7.2

I have just installed it and I confirm an important point: a 34% drop in cpu load. :smiley:

1 Like

I can confirm nearly a 40-50% drop in my CPU usage. What is strange, is I wasn’t getting these results right away. Then about 10 minutes later, it “kicked” in. Whatever the reason is, I’m happy now! :slight_smile: Totally awesome sounds!

1 Like

Lots of interesting sounds but a lot to throw away!
Advantages and disadvantages of open source: rigorous users who take care of the work, but also wacky amateurs (like me :smile:) who botch the work (not like me :wink:) and make things available to be forgotten immediately.
So a lot of work to sort through the mountain of offerings.


This is the same issue with IK ToneX. Which is why I’m happy to pay a modest amount to an individual creating outstanding captures (Amalgam Audio). He has free samples via IK ToneNET. I don’t even bother trying other random user captures any longer.


Yeah, basically the same here - to me, ToneX is just a great sounding platform I can purchase a small number of super high quality tones for and the fact there is a whole load of free stuff is just a bonus I might look at and use sometimes, but I’m certainly not sifting through 10,000 captures constantly trying to find “the one”.

I’m baffled by an issue I’m having with loading some speaker IR’s into different plugins and getting totally different sounds. I haven’t noticed this with pro IR’s like Ownhammer, but couple of IR’s that were uploaded to ToneHunt are sounding very different and it makes no sense to my why. Specifically, when loading into MConvolutionEZ, which I believe @rank13 you use as well right? Could you also check please?

The “Fender Twin Spring Reverb_dc.wav” has a MUCH bigger reverb in MConvolutionEZ than loading directly into NAM. Would you mind taking a little time to load it and compare? Thank you in advance!

In the Vox Bruno pack, when loading it into MConvolutionEZ, it actually adds a delay (as in delay effect aka “echo” not latency delay), which I’d assume isn’t supposed to be in a speaker IR. When I listen to the original wav file, it does sound like there is a delay, but NAM doesn’t play it with a delay. MConvolutionEZ does. What is going on here?

VOX Bruno TB18C1 (1.4 MB)

EDIT: MYSTERY SOLVED! The guys uploading their own IR’s left the length of the IR waaaaaay toooo loooong. 2000ms! Mind you, Ownhammer only offers 200ms and 500ms IR’s. Even the 500 is considered more than you need as almost always, the signal doesn’t last beyond 200ms. So imagine 2000ms!? So once I cut the length of the IR file to sub 500ms, it sounded exactly the same. So apparently, the NAM plugin automatically must only process the first 200 or so ms, that’s why it didn’t play the delay. Very interesting!


Well, It’s the same person

Ok, thank you.

Cool to see he is now starting to create captures in NAM!


Yes, I have just tested his Fender Blonde Bassman 1964 model which is excellent.
The dynamics tracking with the guitar volume control is amazing.


Best one from him IMO. :slight_smile:

Really very nice to use on the bass. The setup I normally favor has some inter-modulation when playing polyphonic on the bass (for example c# + g at position 11/12). This model has no such problems. The only drawback is that NAM takes 23% cpu, where the other setup takes 10%…

You’re using it with a bass with all that distortion? Interesting! Yes it does use more CPU, but as long as I’m not getting pops and clicks I’ll sacrifice summer CPU for better sounds! :slight_smile:

1 Like

I don’t feed it with much input. Then it is clean. After the NAM plugin I placed a low-pass 5KHz 12dB slope, otherwise it glitches a bit in the high.

After all it is a Fender bassman :grinning:

Yes it is, and for a clean capture I wouldn’t even have asked, but in this capture he captured it with quite a bit of distortion. But yeah, you’re right, backing off the input does clean it up nicely!