Where did he say that? And yes I do, more than you care to know.
The power supply in a tube amp isn't a fixed voltage source. It sags, recovers, and modulates in response to your playing; and also your environment. For example in the UK, the mains voltage is 230volts, but we quite often will get more. I've been in my amp techs workshop where he's measured 249 before. That is quite a difference! It varies throughout the day.
Now given that, we know that an amps power supply is not a fixed voltage in the real world. This means an accurate model of it would not be a fixed floating point source either. Rather, you'd model the sag and voltage and other characteristics. DSP guys already know this, which is why amp modelling has come so far in the last 15 or so years. They've built systems to capture these behaviours.
Now. In a DSP system where resources are finite, you have to manage your processing, optimize code, and perhaps even make compromises. For example, instead of running a per-sample tanh function, which can get expensive, you can instead have a look-up table that you index instead. This is far less costly.
You may also run internal oversampling only where it matters. You may also share or reuse state variables between blocks, so that models are interactive without burning CPU.
But if you have an engine and DSP platform where you have more resources to work with, you can go further. You can handle sample-accurate communication, which is exactly what Eric means when he says:
so we can sample accurately capture the interaction between signal and power supply.
Whether it is between the signal and power supply, or the amp and cab block, is kind of irrelevant to my overarching point.
My overarching point is that sample-accurate is a term in DSP that has a specific implication. It isn't just a buzzword, it has a specific meaning:that data is exchanged and processed on a per-sample basis, not in blocks. In other words, there's no buffer-level latency between one part of the model or another.
That's a massive deal for accuracy. It removes the coarsening effect of block-based updates, where interactions only sync every
n samples.
With sample accurate processing, you get tight, real-time feedback loops that make the model feel and response more like actual hardware.
My hunch is that Fractal have been doing this for years already, and that Helix probably wasn't. Until soon
