FM9 sounds different than FM3?

I tried swapping back and forth between my FM3 and FM9 with the same preset. I do hear a small difference. The FM9 does sound a little more clear maybe.
 
Maybe different oversampling among the units?
I've always wondered how an fm3 (which is 25% of an axe fx iii cpu-wise) could run an amp block at the same quality as the ax3 which can only run two.

Maybe someone of you can do an oversampling test to compare the units?
If you don't know how I can explain it.

I'm definitely gonna do it by myself once I receive my fm9.. but since I'm in Europe this could be months away
 
I've always wondered how an fm3 (which is 25% of an axe fx iii cpu-wise) could run an amp block at the same quality as the ax3 which can only run two.

It is a bit counter-intuitive, but the TL;DR is that it's because the sampling rate is independent of the speed of the processor. Since the sample rate is fixed at 48KHz on all three units, there's a minimum CPU speed necessary to run modeling algorithms, which all 3 units exceed.

Even at say, 4X over-sampling, that's 192 KHz. Which means that these modelers need to process 192,000 audio samples (which are just raw numbers representing a slice of waveform) per second. Let's say the FM3 has a (conservative) 800MHz CPU (I actually have no idea what the speed of the FM3's processors are). That's almost 4200x the speed necessary to process audio in real time.

At 1x sampling (48KHz), an 800MHz CPU is almost 17,000 times faster than the sample rate. And that's for the low-end model (the FM3). CPUs are so much faster than the audio sampling rates they're acting on that you could probably run the Fractal Amp Block on a phone processor.

The bottom line is that CPU speed just doesn't scale with quality. Once you're able to run the modeling algorithm, a faster processor doesn't change anything. It just allows you to run more processing per sample. Which is why all 3 units have the same audio quality (both from my own testing and Fractal's statements on the matter). The more powerful units just allow you to do more things at the same time (more blocks per preset, full-res IRs, etc).
 
It is a bit counter-intuitive, but the TL;DR is that it's because the sampling rate is independent of the speed of the processor. Since the sample rate is fixed at 48KHz on all three units, there's a minimum CPU speed necessary to run modeling algorithms, which all 3 units exceed.

Even at say, 4X over-sampling, that's 192 KHz. Which means that these modelers need to process 192,000 audio samples (which are just raw numbers representing a slice of waveform) per second. Let's say the FM3 has a (conservative) 800MHz CPU (I actually have no idea what the speed of the FM3's processors are). That's almost 4200x the speed necessary to process audio in real time.

At 1x sampling (48KHz), an 800MHz CPU is almost 17,000 times faster than the sample rate. And that's for the low-end model (the FM3). CPUs are so much faster than the audio sampling rates they're acting on that you could probably run the Fractal Amp Block on a phone processor.

The bottom line is that CPU speed just doesn't scale with quality. Once you're able to run the modeling algorithm, a faster processor doesn't change anything. It just allows you to run more processing per sample. Which is why all 3 units have the same audio quality (both from my own testing and Fractal's statements on the matter). The more powerful units just allow you to do more things at the same time (more blocks per preset, full-res IRs, etc).
I'm no expert but I don't think it works that way, the processor clock has little to do with the sample rate (especially when complex signal processing is involved) and little to do with the effective speed or power of a cpu.
I think FLOPS (Floating Point Operations Per Second) is a better representation of a cpu's capability.

Furthermore an amp simulation is not just a sample rate converter, it requires a lot of calculations of complex functions, so I definitely don't think it could run on any processor that has a clock frequency equal or greater than the used sample rate (as you seem to imply).

Let's assume a given algorithm (i.e. an amp block) at a given sample rate needs 2 GFLOPS to process the signal within a given latency.
Now, at least for me, it's pretty obvious that if you have the same algorithm and same buffer size (aka latency) but half the sample rate the required number of operations would be cut in half too (=1 GFLOPS), since you're basically feeding it half the amount of data to process in the same time interval.

This agrees with my experience with vst plugins in a daw for example, or with how the axe fx 2 worked (when running 2 amp blocks the oversampling was halved in that device).

Another way to save cpu could be increasing the internal buffer size, thus increasing latency to give the cpu more "time headroom" to process the signal (this agrees with how vts plugins work too, doubling the buffer size cuts in half the daw cpu meter and basically doubles the number of plugins you can run) and the different recording latencies shown in this thread might be an indication of that.

The only way to know for sure is to test and compare the aliasing performance
 
Last edited:
Furthermore an amp simulation is not just a sample rate converter, it requires a lot of calculations of complex functions, so I definitely don't think it could run on any processor that has a clock frequency equal or greater than the used sample rate (as you seem to imply).

I'm not implying that. I was saying that as long as a processor meets the minimum required speed to run the algorithm in real time for a given sample rate, the modeling itself does not benefit from a faster processor.
 
I'm not implying that. I was saying that as long as a processor meets the minimum required speed to run the algorithm in real time for a given sample rate, the modeling itself does not benefit from a faster processor.
Sure, but if the sample rate is changed then the minimum required speed changes too
 
Sure, but if the sample rate is changed then the minimum required speed changes too
Agreed. Which is why I'm saying that since the sample rate is fixed (48K) there's a minimum processor speed necessary for the amp modeling. Even taking into account over-sampling, of which there's a maximum value. So basically once you can run the amp modeling at the highest oversampling rate, which the FM3 can, you don't get higher quality from having a faster processor. You just get the ability to run more things (more blocks, a second amp, etc). Which is why I'm of the opinion that the FM3, FM9, and AxeFX all are equal when it comes to quality. You don't get better sounding amp models by moving from FM3 -> FM9 -> AxeFX.
 
Agreed. Which is why I'm saying that since the sample rate is fixed (48K) there's a minimum processor speed necessary for the amp modeling. Even taking into account over-sampling, of which there's a maximum value. So basically once you can run the amp modeling at the highest oversampling rate, which the FM3 can, you don't get higher quality from having a faster processor. You just get the ability to run more things (more blocks, a second amp, etc). Which is why I'm of the opinion that the FM3, FM9, and AxeFX all are equal when it comes to quality. You don't get better sounding amp models by moving from FM3 -> FM9 -> AxeFX.
You would be right... if the only thing an amp modeler had to do was load samples. But there's a whole lot of math involved in calculating what value the next sample should be. And that's lots of CPU cycles for a single sample.

Bottom line: a faster CPU can give you the power you need for more accurate modeling. But all Fractal products have enough horsepower to run Fractal's wickedly-accurate modeling.
 
You would be right... if the only thing an amp modeler had to do was load samples. But there's a whole lot of math involved in calculating what value the next sample should be. And that's lots of CPU cycles for a single sample.

Bottom line: a faster CPU can give you the power you need for more accurate modeling. But all Fractal products have enough horsepower to run Fractal's wickedly-accurate modeling.
I think we're saying the same thing with just different words 🙃
 
Even taking into account over-sampling, of which there's a maximum value.
There's no maximum value for oversampling, there are only reasonable values for a certain application.
Cliff in his "charts and graphs" thread said that, according to him, the maximum acceptable level of aliasing is -60dB (iirc) with the main signal being 0dB, cuz that's the threshold of audibility.
But the aliasing level also depends on the amount of distortion, e.g. a clean amp like a band commander fed with a single coil signal would probably produce aliasing at -60db with just 2x oversampling, while a high-gain monster like the skull crusher with all controls at max played with super-high output pickups would probably need 64x or even 128x.
So I assume Cliff just took the real-world-worst-case scenario when doing his tests and deciding what oversampling to use in the axe fx III, but that doesn't mean the fm3 or fm9 are able to run at the same oversampling rate and maybe a compromise was inevitable.

which the FM3 can
How do you know that? That's the point
 
Last edited:
There's no maximum value for oversampling, there are only reasonable values for a certain application.
Cliff in his "charts and graphs" thread said that, according to him, the maximum acceptable level of aliasing is -60dB (iirc) with the main signal being 0dB, cuz that's the threshold of audibility.
But the aliasing level also depends on the amount of distortion, e.g. a clean amp like a band commander fed with a single coil signal would probably produce aliasing at -60db with just 2x oversampling, while a high-gain monster like the skull crusher with all controls at max played with super-high output pickups would probably need 64x or even 128x.
So I assume Cliff just took the real-world-worst-case scenario when doing his tests and deciding what oversampling to use in the axe fx III, but that doesn't mean the fm3 or fm9 are able to run at the same oversampling rate and maybe a compromise was inevitable.


How do you know that? That's the point
You are right about oversampling
I have AxeFx3, FM9, FM3 all three units and done sine sweep aliasing level test, Axefx3 have scary low level aliasing noise when only 1 amp block,
the oversampling level (AxeFx3 1 amp block)>(AxeFx3 2 amp block)>(FM9 1=2 amp block=FM3 1 amp block)
 
You are right about oversampling
I have AxeFx3, FM9, FM3 all three units and done sine sweep aliasing level test, Axefx3 have scary low level aliasing noise when only 1 amp block,
the oversampling level (AxeFx3 1 amp block)>(AxeFx3 2 amp block)>(FM9 1=2 amp block=FM3 1 amp block)
Interesting, would you mind sharing some graphs of those tests?

PS: this bit in particular is quite interesting
(AxeFx3 1 amp block)>(AxeFx3 2 amp block)
IIRC when the axe fx III came out it was stated that, unlike the axe fx II, the quality wasn't reduced when running 2 amp blocks. Maybe that changed over time with firmware updates.
I'll verify that asap
 
Last edited:
Interesting, would you mind sharing some graphs of those tests?

PS: this bit in particular is quite interesting

IIRC when the axe fx III came out it was stated that, unlike the axe fx II, the quality wasn't reduced when running 2 amp blocks. Maybe that changed over time with firmware updates.
I'll verify that asap
I think the quality wasn't reduced, It's even higher, when only 1 amp block the aliasing noise nearly inaudible when amp gain set to normal high gain level
 
I think the quality wasn't reduced, It's even higher, when only 1 amp block the aliasing noise nearly inaudible when amp gain set to normal high gain level
Yeah, I was speaking in relative terms. I just meant the quality is lower when running 2 amps compared to just 1 amp.
 
There's no maximum value for oversampling, there are only reasonable values for a certain application.

You make good points. I had some misunderstanding about how oversampling works.

How do you know that? That's the point

From the numerous statements by Cliff and Fractal staff that the modeling quality is the same between the three units.

At this point, it's all speculation though, isn't it? We don't actually know the oversampling characteristics of Fractal's amp modeling, do we? We just have the aliasing performance that can be measured.

At the end of the day, what my ears tell me is that there's no difference between the units. Which is good enough for me.
 
I think the difference comes down to the variable impedance that the FM3 lacks, It's most noticeable when you try various style pickups such as low output vintage alnico. Not so much with higher output ceramic, maybe it's me.
Unless you're using one of the special blocks that takes advantage of the Auto setting or you're manually setting the the impedance lower then that shouldn't matter as far as I'm aware, since both units should have a 1M Ohm impedance in the case.
 
(AxeFx3 1 amp block)>(AxeFx3 2 amp block)
Confirmed.

This is aliasing with one amp block:
Immagine 2021-11-05 140242.png

This is with 2 amp blocks (only one connected):
Immagine 2021-11-05 140201.png

Roughly a 12dB difference in aliasing noise

PS: if anyone wants to repeat this test on the fm3 or fm9, just load the attached preset, record 30 seconds of audio in your daw and send it to me
 

Attachments

  • aliasing test.syx
    48.2 KB · Views: 2
Last edited:
Well, I can tell you my personal findings. When my FM3 arrived, default stock presets just didn`t sound good for me. Which was why I went for the Cygnus Live Gold from AB. Even after I removed AustinBuddys awesome Cygnus Live Gold and reloaded the latest official Cygnus presets from Fractal back on the FM3 to get it ready for Sale. Same thing. Blows the mind tbh.

But booting up the new FM9 here just now, the default presets that came with the FM9 do sound better right outta the gate. At least to my ears. Kinda shocked because I was expecting the stock FM9 presets would turn me off like the FM3s did. Go figure. I`m still going AustinBuddys Cygnus Live Gold for my FM9.... but Fractals FM9 stock presets.... all the 1s I tried are pretty sweet and workable/usable too.
 
Back
Top Bottom