NAM support announced by Fractal Audio

[Fireproof]

No, developers building with NAM don't need to update/sync changes unless they feel it important to do so.
They can either fork the NAM core repository, or just download the source code as a zip, then build products as they please.
So, it's not a big deal at all.

The NAM file format (ie: what you need to be compatible with) hasn't had breaking changes in years.

Thanks for sharing. Like I said - I really don’t know much about the tech so that was a helpful correction.

 

[ChordInversion]

I take the totality of your response and get the overall sentiment, I think we (and a lot of people) mostly agree on 8/10 things.

Cool! Honestly, if any two people agree on everything, one of them may be unnecessary, and I'd presume that one was me.

On your point above... Adobe Photoshop recently upped the pricing of the lightroom+photoshop bundle +52% $$$ / month. Ontop of that they keep adding AI features that will start to use credits, so if you're a power user you'll be up for even more when you exhaust your allotment per month. At the very least these features are crazy powerful and pure insanity for the photo editing world, but the writing is on the wall with all this stuff, higher fees, more AI (yuck)

No argument there. Yuck indeed. We better start finding and making alternatives to the big tech stuff soon, because they're coming for us with their intent to leverage their opligopolies/de facto monopolies. The myth of infinite meteoric growth and acceptable stratospheric PE ratios is finally losing its legs.

One last lol, they can finally move to Fractal when profiling is on an equal sized unit

Hah! Well played.

 

[ChordInversion]

Saying capture tech is table stakes is stating one’s view of the market place, not a demand and not belittling those who don’t want it.

OK, but I was also stating my own view, and I was in no way belittling people who want capture. By all means quote where I did if you disagree. Unless you were speaking of others, of course.

I believe that the only two players that MIGHT have survived without it were Line 6 and Fractal, but it was starting to seem like they would both risk losing sales without it. Prior to Cliff’s announcement, people were actively talking about selling Fractal gear and going to Stadium, and Proxy is still a complete unknown!!!

The fact that they both announced upcoming support within 6 months of each other, suggests that they both think it’s an important feature from a competitive point of view. I get that some people don’t want or need it, but like it or not, capture support IS becoming like IR support. It doesn’t matter how good your built in cabs are, almost all the players support third party IR’s and a large portion of the customer base simply expects that. That’s the reality for capture tech in 2026 or 2027 at the latest.

Yes, I definitely lost on this one, as I have with other ones in the past.

If you want more bad news, I honestly think component modeling is a dying tech. It won’t go away, but it will be a shrinking niche. I see Line 6 and Fractal keeping it front and center, but for smaller and low budget players, the time and cost to develop the tech and the models is going to be harder and harder to justify when machine learning can make profiles/captures so easily.

No problem. I can stick with what modelers I have or go back to amps.

The idea that the tone controls must behave exactly like the real amp is a huge red herring. Most new buyers will have no idea how the controls on the real amp react, and most of the rest will only be familiar with a small number of real amps. They need tools to take a capture that is close to what they need and make small adjustments, and that exists already. It works far far better than a lot of people realize, so no, you really don’t need to find THE perfect capture, you just need one in the ball park. For the vast majority, getting a good tone will be what matters. Plus, for those who do care, capture tech is going to solve that “problem” as well anyway with dynamic captures or whatever you want to call it. Again, a little AI programming and a few servos to twist pots will be able to capture the full range of an amps tonal possibilities in a fraction of the time and at a fraction of the cost of adding a new amp model. That’s not 10 years from now, it’s something people are working on today.

Sure, belittling people who don't share your preferences is definitely not cool. Telling people what they do and don't need would be belittling, and that would be wrong. Totally.

Anyway, my reason for liking modeling speaks to a whole different use case than simply "find a good sound in the ballpark." I can get into it in another thread some time, because I don't want to come off as saying "my way is the good way." It's just my way. And my way doesn't work with captures. I'm sure you won't insist that my way must work with captures, or that my way is stupid, because that would be belittling.

 

[AlbertA]

But (and I don’t know much about the underlying code), building your device on open source code seems like it would introduce some other challenges. Now, not only do you have to focus on your own tech bug fixes and feature enhancements, but as the open-source code changes, won’t you have to now keep up with that. Being dependent on ever evolving open-source code seems like it might come with its own share of challenges…. But maybe it’s not that big of a deal??

You don't have to use the code as is - think of the NAM Core repo as the reference implementation.

The important part is the topology, specially the more popular topologies being used by the people uploading captures. The main one being WaveNet. And further, the configuration for that topology (layers, channels, dilations, kernel size, activation function)

This is where the distinction between STD and xSTD comes from - both are WaveNet architectures, but xSTD restructures the topology for more temporal depth and smoother dilation spacing, using more layer groups with larger kernels and ternary dilations.

That configuration is specified in the NAM file, which is a JSON based container that stores the architecture type, the architecture configuration, the full set of trained weights and some other metadata.

So then an implementor has some options:

• Use the reference implementation as is
• Fork the reference implementation
• Reimplement an engine that essentially replicates what the reference implementation does, maybe more efficiently, maybe more adapted to the processor architecture of your choice
• Implement your own neural network architecture - perhaps also based on WaveNet - this is probably what makes most sense for most manufacturers as you decouple from any changes to NAM itself.
  • This is the approach SonicCake,Hotone,Valeton are using
  • Here since we have the NAM reference implementation, you can run tests signals say with the NAM core processing code, which then serve as the "teacher" or reference signals to match
  • You then train your own proprietary architecture using those reference outputs
  • Voila! you know can support any standard NAM file since you are converting to your proprietary neural network architecture.
  • For example, IK Multimedia could decide to support NAM this way with ToneX
  • Line6 and proxy could do things like this too (and since there will be connection to the cloud, training can be more extensive).
  • Fidelity loss (if any) will mainly be dictated by the proprietary architecture you chose (and some to the test signals you choose to use).

The NAM file format (ie: what you need to be compatible with) hasn't had breaking changes in years.

Well it's a little bit more complicated than that - see above.

 

[AlbertA]

I don’t get what that comment meant though because if Cliff puts a big screen with touchscreen with knobs under it, it’s gonna look like a game boy as well so when he says others will look like a game boy it seems like he’s putting others down butI guess I just don’t understand what he means

That's because the "game boy" comment had nothing to do with looks. The context was processing power.

Fractal used hyperbole to essentially say: "Our next gen products are so powerful compared to current or competitor devices that it dwarfs them the way a modern supercomputer would dwarf a 1990s handheld."

 

[Mike Oliphant]

Well it's a little bit more complicated than that - see above.

My point was meant to counter the sentiment I've seen that NAM is some fast-moving target that you would be crazy to support in your hardware because those open source rascals will just pull the rug out from underneath you.

If you had a NAM implementation (either using NAM Core, or your own custom version) back in early 2023, it would still work with all existing NAM models today - including "xSTD" models if you have the processing power.

 

[paisleywookiee]

But I still like my Gameboy.

 

[timbuck3]

That's because the "game boy" comment had nothing to do with looks. The context was processing power.

Fractal used hyperbole to essentially say: "Our upcoming technology is so powerful compared to current or competitor devices that it dwarfs them the way a modern supercomputer would dwarf a 1990s handheld."

 

[timbuck3]

Keep huffing copium!!! Ok?

 

[laxu]

That's because the "game boy" comment had nothing to do with looks. The context was processing power.

Fractal used hyperbole to essentially say: "Our next gen products are so powerful compared to current or competitor devices that it dwarfs them the way a modern supercomputer would dwarf a 1990s handheld."

Which begs the question: "What is Fractal going to do with all that power?"

Most people never max out their Axe-Fx 3 or FM9. So just throwing more processing power at it doesn't mean anything when majority of users will be running fairly standard signal chains that might just barely not fit a FM3 if you go for the kitchen sink approach.

Axe-Fx IV is likely using the TI C7x architecture (Axe-Fx 3 is C66x). One key feature of C7x is that is has accelerators suitable for Neural Networks, which is likely what is used for NAM support in the same way as FIR/IIR accelerators are used for IRs. It'll be interesting to see if Fractal finds NNs to be useful for component modeling too.

What about "FM4" and "FM10"? If Fractal sticks with Analog Devices, we'll have to see what they do for NAM support on those. ADSP does not seem to have NN accelerators specifically on the ADSP-SC594 or SC598, so maybe Axe-Fx IV can run a pile of NAM captures at once, and the floor units just one or two blocks at once, maybe with channels to switch between captures?

Still, to me NAM support is about the least exciting feature Fractal could announce.

 

[Orvillain]

Still, to me NAM support is about the least exciting feature Fractal could announce.

I guess the real question is ... what other features do we want?

And I do mean features. Not workflow or usability enhancements.

What raw features are we really missing from Helix Stadium (let's just jump straight there) and Axe III at this point?

 

[[Nathan]]

Still, to me NAM support is about the least exciting feature Fractal could announce.

With stadium and QC having both it would have just left a bit of a gap in peoples purchasing decisions, now that its coming its just levelled that side of the playing field (whether you use it or not). So yeah its not ultra exciting at first glance but it shuts the door on the conversation of people "needing" 2 units to achieve those kinds of things... Fractal supports profiling, yeah sick.

I guess the real question is ... what other features do we want?
And I do mean features. Not workflow or usability enhancements.
What raw features are we really missing from Helix Stadium (let's just jump straight there) and Axe III at this point?

I wouldnt even know, its all there and more for my uses. Maybe the ability for people to create chase bliss style effects with FX building blocks and then save it down to a convenient block preset for us mere mortals to mess with as end users. I know theres kitchen sink presets already but being able to map that stuff to a couple of knobs or modifiers in a single block would make usability a lot easier. Maybe its too hard I dunno... I DUNNO... just keep making the modelling algos better so I can play a 5153 endlessly.

 

[Dimiguitar]

I guess the real question is ... what other features do we want?

And I do mean features. Not workflow or usability enhancements.

What raw features are we really missing from Helix Stadium (let's just jump straight there) and Axe III at this point?

Probably you don't count it as a feature, but stadium itself lacks 734584237598472984 amp sims compared to fractal. But sure, in terms of what more guitarists could ask for, after a given threshold, I don't know.

 

[MirrorProfiles]

Which begs the question: "What is Fractal going to do with all that power?"

Most people never max out their Axe-Fx 3 or FM9. So just throwing more processing power at it doesn't mean anything when majority of users will be running fairly standard signal chains that might just barely not fit a FM3 if you go for the kitchen sink approach.

Axe-Fx IV is likely using the TI C7x architecture (Axe-Fx 3 is C66x). One key feature of C7x is that is has accelerators suitable for Neural Networks, which is likely what is used for NAM support in the same way as FIR/IIR accelerators are used for IRs. It'll be interesting to see if Fractal finds NNs to be useful for component modeling too.

What about "FM4" and "FM10"? If Fractal sticks with Analog Devices, we'll have to see what they do for NAM support on those. ADSP does not seem to have NN accelerators specifically on the ADSP-SC594 or SC598, so maybe Axe-Fx IV can run a pile of NAM captures at once, and the floor units just one or two blocks at once, maybe with channels to switch between captures?

Still, to me NAM support is about the least exciting feature Fractal could announce.

I agree - I'd rather see the extra CPU used for things like being able to route mono signals more freely and easily without needing to make panning blocks and doing weird splits. Similarly, being able to use all digital I/O at the same time rather than being limited to only one type.

There's also things like running 2 presets at once (say you had 2 guitarists or guitar and bass with their own chains, and they could each switch internally on one global preset). Or perhaps (as Nathan said) combining several blocks into a single "macro" block for the use. And a big one would be to finally consolidate amp channels and modes into single amp models because IMO that really does make things feel a bit dated compared to having an all in one model.

As much as I love modelling accuracy kicking on and getting better, it would be a shame if thats the only area that gets any attention with more CPU available. I highly doubt that'll be the case, although I'm sure there are likely a few ideas Cliff has that need the extra grunt.

 

[Los-D]

Would be interesting to see if the NAM block could also somehow be connected/fed to a Tone Match block. So you could use a NAM file as reference, but still keep all the modeling flexibility. Probably easier than full parametric NAM(?).

Total speculation, of course.

(In this case, you would play using a tone match block, not a NAM block. Just thinking about using the NAM as an input)

 

[laxu]

I guess the real question is ... what other features do we want?

And I do mean features. Not workflow or usability enhancements.

What raw features are we really missing from Helix Stadium (let's just jump straight there) and Axe III at this point?

• Room mics for cab sims. Yes, room reverbs will do most of the same job but it's not 100% the same thing.
• More experimental effects. They're not for me, but there's a subset of users that enjoy messing with them.
• Full blown synth blocks that allow you to control them e.g via MIDI notes, or MIDI notes generated from your guitar. This would allow e.g drone notes or whatever to be triggered via footswitching.
• Cab positioning in 3D space for headphones use ala Boss Waza Air.
• Multiple loopers in different parts of the signal chain?

Honestly 99% of what I want from next gen modelers is workflow and usability stuff.

 

[LP59]

I guess the real question is ... what other features do we want?

And I do mean features. Not workflow or usability enhancements.

What raw features are we really missing from Helix Stadium (let's just jump straight there) and Axe III at this point?

One thing comes to mind right away: when you switch channels or presets, load and run both the old and new channel/preset and cross fade to the new one. That would require a lot more cpu. Kemper does this and it is very useful. Also, reduced latency and aliasing could be accomplished with additional cpu power.

But, in general, I agree with what I think you're saying. More cpu power in the IV would be nice, but it's the workflow issues that really need attention, so I'm not sure the cpu power is the right thing to tout at this point.

 

[deathbyguitar]

I guess the real question is ... what other features do we want?

And I do mean features. Not workflow or usability enhancements.

What raw features are we really missing from Helix Stadium (let's just jump straight there) and Axe III at this point?

I'd love to see :

1. Independently controllable signal paths via multiple FASLINK connectors.

2. More processing power to allow 3 or 4 amp blocks at once and additional effects at one time. That's probably coming.

3. Slightly expanded mixer functionality and I/O aimed at people on IEMS so they don't need a separate mixer. If you're already using the unit for guitar and bass, it wouldn't be that much more complicated to pipe in a mic or two, process them with comp and EQ, then have 2 different outputs for 2 different mixes (guitars+bass+vocals) going to 2 different IEM transmitters. Maybe you have one mix set to prioritize the singer's levels, another for a general purpose mix. This is how I'm currently running my band with the Axe-FX but it'd be nice if I could ditch the Midas MR12 mixer and have all that stuff in the Axe.

Overall, I just wanna see more features that empower smaller artists to do more with less. Maybe that kind of thing would add a bit cost to the Fractal but you'd be able to save money and space from omitting a mixer. A boy can dream....

 

[AlbertA]

What about "FM4" and "FM10"? If Fractal sticks with Analog Devices, we'll have to see what they do for NAM support on those. ADSP does not seem to have NN accelerators specifically on the ADSP-SC594 or SC598, so maybe Axe-Fx IV can run a pile of NAM captures at once, and the floor units just one or two blocks at once, maybe with channels to switch between captures?

I imagine they would like to keep the same family of SoCs this time around to decrease development burden. The last time I looked at the Jacinto 7 family, they offered SoCs at different price points which one could theorize could map to next-gen devices at different price points.

Post in thread 'Next Gen Fractal Speculation Thread'

Jun 13, 2025

So taking a hint based on Fractal Audio's careers page, looks like they will use TI's Jacinto7 family for their new platform

There's some newish SoC's in that line that are stupidly powerful

TDA4XX line - 4 C7x DSP cores, 8 ARM Cortex A72 cores around the $180 range,
TDA4AEN-Q1 - 4 ARM Cortex A53 cores, 2 C7x DSPS) around $60 range
AM67 - 4 ARM Cortex A53 cores, 2 C7x DSPs around $50 range
AM62A3-Q1 - 4 ARM Cortex A53 cores, 1 C7x DSPs around $30 range

TDA4VH-Q1 data sheet, product information and support | TI.com

TI’s TDA4VH-Q1 is a SoC with Octal Arm® Cortex®-A72, 32 TOPS of AI, C7xDSP, and GPU for vision perception and analytics. Find parameters, ordering and quality information

www.ti.com

TDA4AEN-Q1 data sheet, product information and support | TI.com

TI’s TDA4AEN-Q1 is a SoC with Quad Arm® Cortex®-A53, 4 TOPS of AI, and C7xDSP for vision perception and analytics. Find parameters, ordering and quality information

www.ti.com

AM67 data sheet, product information and support | TI.com

TI’s AM67 is a Arm®Cortex®-A53 SoC with triple display, 3D graphics, PCIe 3, USB3, 4K video codec for HMI. Find parameters, ordering and quality information

www.ti.com

...

• AlbertA

• 

 

[laxu]

I'd love to see :

1. Independently controllable signal paths via multiple FASLINK connectors.

That's a neat idea. Would not necessarily even need multiple, but just chaining two FC units together and allowing the footswitching to be assigned to manage different signal paths. Goes with that "two presets at once" idea for sure.

But I think multiple guitarists or guitarist/bassist using the same Axe-Fx are quite few so it'd be a pretty niche feature.

3. Slightly expanded mixer functionality and I/O aimed at people on IEMS so they don't need a separate mixer. If you're already using the unit for guitar and bass, it wouldn't be that much more complicated to pipe in a mic or two, process them with comp and EQ, then have 2 different outputs for 2 different mixes (guitars+bass+vocals) going to 2 different IEM transmitters. Maybe you have one mix set to prioritize the singer's levels, another for a general purpose mix. This is how I'm currently running my band with the Axe-FX but it'd be nice if I could ditch the Midas MR12 mixer and have all that stuff in the Axe.

Overall, I just wanna see more features that empower smaller artists to do more with less. Maybe that kind of thing would add a bit cost to the Fractal but you'd be able to save money and space from omitting a mixer. A boy can dream....

I can agree with that. The current mixer block sucks because instead of being based on inputs/outputs, it's based on rows which is quite unintuitive.

I'd like to see a node-based connection system in the next gen, even if it aligns those nodes on a grid. Get rid of the "shunt" blocks.