Amp modeling progress

From Axe-Fx II Wiki

Jump to: navigation, search

The information on this page supplements the official manuals.


H amps.png

About amp modeling

This page documents the progress of Fractal Audio's amp modeling, according to firmware release notes.

"The hardest part of modeling an amp is getting the various controls to match the actual amp. If you don't care if the tone, drive, etc. controls behave the same it's much easier as we have software that learns the input EQ, output EQ and gain. The problem is then people go "the model doesn't sound the same as my amp if I turn the drive all the way up and bass all the way down". So to accurately model the control behavior we need a schematic and the actual amp (as the schematics often don't indicate the pot tapers). Truth is amps are more similar than people think. You can make almost any high gain amp sound like any other high gain amp with a few EQ tweaks which is basically what the designers do. For example a Bogner is basically a boosted Marshall with a different treble pot taper. Another popular new amp is basically just a JTM45 clone with a couple minor changes. In fact the schematic I got from the designer was a JTM45 schematic with markups. The scary thing I've learned is that a lot of these amp "designers" don't really even understand what they are doing. They don't have degrees in engineering and lack even basic circuit theory. They take existing designs and tinker with them changing circuit values. The basic topology of the amps are unchanged. So many of these new amps are nothing more than clones of old designs with some minor changes. Things you can do in the Axe-Fx with all the EQ options available. There are only a handful of guys that really understand circuit theory and know what they're doing: Alan Phillips from Carol-Ann, Stevie Fryette, John Suhr, and several others. The vast majority are glorified technicians that are just making clones of existing designs with minor modifications. A good example is the Marshall 18W. There are numerous clones and amps inspired by this design. The problem is that the original design is flawed. You can make that amp sound much better with some minor changes to the phase inverter (or grid stoppers) but none of these amps do that. They all use the same PI design which overdrives the snot out of the power tubes making the amp shift into Class-B operation resulting in fizz and crackling on the decay." source


Multi-Point Iterative Matching and Impedance Correction (MIMIC) is a technology that identifies deviations in the response of the simulated amplifier to the actual amplifier and generates corrective data bringing a level of accuracy that has been heretofore unachievable. MIMIC was introduced in firmware 10.

MIMIC whitepaper (PDF).

If Power Amp Modeling is off, the pertinent aspects of MIMIC are defeated. source

"The whole impetus for V10 was actually due to the Rectos. I hooked up my Dual Recto one day and was A/B'ing to the Axe-Fx. The real amp had this fizz and sizzle on the notes whereas the Axe-Fx did not. If you play a single note you can hear a crackling noise on the decay, almost like frying bacon or something. The Axe-Fx was more of a smooth fuzz on the decay. Outside of a mix context that sizzle can be a bit annoying but in a mix it cuts like mad. Anyways I sat there wondering why the Axe-Fx lacked this sizzle. No amount of tone matching would recreate it. So I wrote a statistical analysis module (it's hidden in the TM block btw) that allowed me to compare the distortion profile of the real amp to the model (there's a hidden switch that gathers statistics along with tone match data). I was a bit shocked at the results. The real amp had a much harder clipping profile than the model. So I tested a bunch more amps and found the same thing. Then I decided that the best thing would be to simply compare the statistics and then have a feedback loop that adjusts the parameters until the statistics match. This, along with the other stuff in the whitepaper, was the foundation of MIMIC."
"I wrote a statistical analysis module (it's hidden in the TM block btw) that allowed me to compare the distortion profile of the real amp to the model (there's a hidden switch that gathers statistics along with tone match data). I was a bit shocked at the results. The real amp had a much harder clipping profile than the model. So I tested a bunch more amps and found the same thing. Then I decided that the best thing would be to simply compare the statistics and then have a feedback loop that adjusts the parameters until the statistics match. This, along with the other stuff in the whitepaper, was the foundation of MIMIC."
"MIMIC is distortion profile and frequency response matching. Hidden in the debug version of the firmware are special test tones and analysis modules that allow me to compare the real amp to the model." source

Firmware 4 and earlier

Amp modeling in the Axe-Fx II was originally called G2 and Virtual Vaccuum Modeling modeling (see Fractal Audio website). Part of it was ported back to Standard/Ultra firmware 11.

"As you listen to clips from modelers what you start to recognize is a certain "stationary" aspect to the tone compared to the every-changing tonality of a tube amp. Another thing is finger response. With a good tube amp you can vary the tone quite a bit just by how you fret the note and attack it. Modelers tend to make every note sound the same. So I tested some hypotheses and came to the conclusion that it's because a real vacuum tube has a transfer function that is not static. The transfer function is dependent on time, frequency and amplitude. Where you really hear it is in the in-between regions where the tube is just starting to distort. At first I tried some dynamic transfer functions but that was a lesson in futility. So then I created the VVT stuff. In VVT there is an actual vacuum-tube replica in software. You enter the values of the resistors and capacitors on the grid, cathode, etc. and it behaves just like a tube complete with Miller effect, cathode memory, etc. The problem is that it requires an obscene amount of horsepower so the only solution was a dedicated DSP. The other big part of the G2 sound is the output transformer modeling. The OT distorts and as it distorts its inductance decreases which changes the bandwidth and loop characteristics." source

Firmware 5

"The new firmware (V5) is all about dynamics. The power amp modeling was totally rewritten. If you listen to those early VH albums you can hear the amps knock and ring when he hits them hard. Lots of work went into discovering why that happens and replicating it. It's due to several factors: the power supply sagging, the bias point shifting as the supply sags, the screens ringing and power compression in the speakers." source
"Much improved grid modeling in Amp block preamp and power amp stages. New modeling very accurately replicates grid conduction and resulting bias excursion. This results in a more dynamic, thicker and bouncier tone. The power tube grid conduction parameters are exposed to the user in the GUI. The Bias Excursion parameter controls how much the grid voltage droops when the grids conduct. The Excursion Time and Recovery Time parameters control the time constants associated with the excursion."
"Added dynamics processing to Amp block. A new tab, “DYN”, in the amp block, allows adjusting various parameters of the dynamics processor along with several other parameters related to amp dynamics. The Dynamics parameter controls the amount of dynamics processing and models the interaction between the power amp, power supply and loudspeaker under high power-level conditions. The Dynamics Time parameter (ADV tab) controls the time constant of the associated processing."

Firmware 6

"Almost everything has been reworked. Almost all the "real" amps are "matched" to the actual amp. This involved much more than just Tone Matching. It also included gain matching, harmonic content matching, tone-stack matching, etc." source

Firmware 7

Firmware 8

Firmware 9

Firmware 10

"The models are statistically matched to the amps in both frequency response and distortion profile. In some cases this gives a certain sizzle. Some may not like this but it is accurate." source
"The amp modeling in V10 is very similar to V9. There are only a few changes to the algorithms. The big difference between V9 and V10 is the inclusion of the MIMIC data and many of the cabs were remixed. Of those two the remixed cabs will be far more audible. So to get that V9 sound the most important thing is to go to the cab block and select the equivalent V9 cab(s). These are indicated by "(V9)" in the name. For example, if you preset is using "1x12 Brit G12H30" it has a remixed cab. Therefore change the cab to "1x12 Brit G12H30 (V9)". source

Firmware 11

Firmware 12

"The harder the virtual power amp is driven the more noticeable the improved preamp modeling will be. Before I was treating the models as separate preamps and power amps as though you were using a rack system with a preamp unit and a power amp. I did some tests and noticed that there is quite a bit of interaction between the power amp and preamp in an integrated amp (combo or head). So now the modeling feeds the power supply voltage from the power amp algorithm back into the preamp algorithm. As the B+ sags (and the screens droop and bounce) the preamp is affected as it will in a real amp. What happens is that as you hold a note or chord the B+ sags. The preamp voltage sags as well but at a slower rate due to all the extra capacitance and resistance between the screen voltage and the preamp tubes. Eventually the preamp voltages sag enough to compress 10-20% depending on the amp. This is a couple dB or so of compression. It's a slow compression though so your pick attack is unaffected. I don't think I hear much, if any, difference but the feel is definitely different. It feels a lot easier to play like I don't need to press on the strings as hard. I imagine the low end will be affected as the compression will cause less drive into the power amp and therefore less saturation at the resonance of the speaker." source

Firmware 13

Firmware 14

Firmware 15

Firmware 16

Firmware 17

Firmware 18 (G3)

"G3 modeling is a proprietary technique that learns the characteristics of a nonlinear circuit." source
"I have three units here in my lab. One has 17.04 and the other two have 18.00. A/B'ing them it's immediately apparent and as Manny noted the most apparent thing is the punch in the chest aspect. That's exactly what I mentioned about a week or two that I noticed being the biggest difference. You can feel the speakers "jump" for lack of a better term. One caveat, though, is that punch is dependent upon MV. The higher the MV the more punch but too high and things can get stuffy and/or raspy. You have to find the sweet spot and that depends on the other controls as well as your guitar and playing style. The higher the MV, the more the power supply sags and bounces. That power supply sag and bounce is partially what contributes to the punch." source
"You should notice better midrange and treble clarity on certain high gain amp models compared to the beta. I did a lot of critical listening and comparing of the new algorithms over the last few weeks and noticed a tiny bit of "signature" to the sound. I traced that to the triode cutoff formula being too "soft". The new formula removes that signature which also yields increased clarity and string separation for models that rely on preamp distortion." source
"In general the improvements are subtle but there should be improvements in the "bounce" and slightly smoother power amp distortion. I realized I wasn't accounting for the fact that bias excursion is a function of the B+ voltage. I also derived more accurate formulas for the behavior of the various power tubes when they enter saturation. The improved OT modeling results in fewer ultrasonic overtones which, while inaudible on their own do result in more intermodulation "garbage" when playing multiple notes so you should hear better string separation and chord clarity." source

Firmware 19

Firmware Quantum 1.x

"Personally I consider Quantum the biggest improvement from previous firmware than any other release. The day I first got it working I was so shocked at how good it sounded I jumped out my chair. I then crafted an email to Matt with the title "I did it!!!". I'm sure he'll be happy to verify." source
"Q1.01 models the power tube plate current slightly more accurately than Q1.00. It's a subtle difference but some may prefer the more idealized sound of Q1.00 so I made it a menu option." source
"Tube amps are amazing in some ways. The just naturally do the right things. One of these things is that the grids sort of "self-align" so that the distortion stays on the edge. You can tell when a modeler isn't modeling this right as you will get fizzy decay (some people erroneously call this "crossover distortion") and flubby bass. If you can't boost treble without it getting harsh then it's not modeling the grids properly. If you need to put a shelving EQ to roll off some bass before the model then it's not modeling properly. With this latest version of Quantum you can put an EQ after the amp or cab and crank the highs as much as you want without it getting harsh and causing ear fatigue." source

Firmware Quantum 2.x

"The previous firmwares were based on theoretical tube models. One thing I began to notice is that the theoretical models simply weren't agreeing with measurements. So after much research we realized that the theoretical models are just that, theoretical and that real tubes do not behave exactly as theory predicts. I can't reveal the details of what aspects don't agree and why those aspects are important. Q2 is based on the actual tube curves which gives a different character to the distortion and a more bouncy and responsive feel." source
"Many of the amp models had their speaker parameters updated based on new measurements." source

Firmware Quantum 3.0

Firmware Quantum 4.x

Firmware Quantum 5.x

Firmware Quantum 6.x

"The preamp modeling in 6.00 is the same as 5.xx except the parameters for the default tube type (12AX7A SYL) are different. The Sylvania 12AX7A is more nonlinear than other 12AX7As which results in more dynamics but will also result in more "background" distortion because the waveform is being distorted even when it isn't being clipped. The JJ version is more linear which will result in a tighter tone and less background distortion but less dynamics. For 6.01 I've also added back the old 12AX7B type which is the most linear of the types and clips hard. People who play with lots of gain tend to like this as it results in tighter tone and more aggressive harmonic content. There are two primary parameters associated with our preamp tube models. "Preamp Hardness" determines how abrupt the tube clips when it enters the saturation region. There is another parameter that determines how nonlinear the tube is between cutoff and saturation. This is currently not exposed to the user but I've been contemplating adding it. I've also changed the default type for British amps to the ECC83 model as these amps typically were equipped with ECC83s (duh). The ECC83 was the European equivalent of the 12AX7A and tended to be a bit more linear and clip a little harder." source

Firmware Quantum 7.x

Firmware Quantum 8.x

Firmware Quantum 9.x

Personal tools
Wiki sections