ToneX Modeling could be faster (technical)

[FenderSim1]

Hello IK Team,
I've recently got myself a new NVIDIA RTX 4070 card and was a little bummed by the capturing performance. Some people on forums claim that a extende capture takes at around 30 minutes on their GTX 1650 which i find hard to believe.

Currently the advanced training takes 45min and 30 seconds on my system. What's odd is that the GPU usage stays at around 30% only so i believe that the performance could be improved greatly by optimizing the usage.

Some assumtions.
- The batch size for the training data is set too low on purpose so that weaker systems don't run out of memory or can even handle the training
- the used pytorch or maybe xformers (if used) are not optimized for NVIDIA cards?

Is there anything else that could limit the GPU usage, because it feels like driving in first gear at a max of 10 mph.

My CPU usage is at 50% while training

[jhall123]

I have an i-7 based Dell Optiplex 9070 with 32 gb systen ram and a 1-TB SSD. It is an AMAZON refurb I got for $279.00. This is a common "sleeper machine that was a good "power desktop" for office work so there are thousands of them available.
I then spent $279.00 on an NVIDIA 1660 6GB based GPU card, $59.00 on a new power supply that can power the new card and $13.00 on a power supply adapter for the motherboard (Dell used a proprietary 8 pin molex connecter). So for $512.25 after tax I have a computer that will create an advanced profile in 16minutes from the time I press "Start Training". I built this computer specifically for this use in my "guitar room". I was so impressed by it I went right back on Amazon and put together a second identical computer for my "studio" (a full living featured cottage behind my house that I decided not to rent as an apartment any more).
It seems I stumbled onto the sweet spot in horsepower and optimization. It will do a default profile in 4 minutes which really sounds the same to me for my big iron Marshalls and blackface Bassman amps. I did a profile with (believe it or not) a Bugera 1960 infinium With the master on 10 and both volumes at 9 with an MXR raijin OD in front at the advanced setting (I go by what I've read about advanced having a slight accuracy advantage over default for high gain but honestly I don't hear a difference if you tweak the treble to7 on the model after creating)
This amp was turned up so loud that the power light was blinking about 3/4 of the way into the capture. But WOW is the lead tone AMAZING! I use the VIR function with the 4x12 green and it is to die for!
As a side I tried a Gforce MX 1030 based card with 2Gb GPU memory to see the difference and it wouldn't start training at advanced but would do a default capture in about 6 minutes.