Reply to thread

Message: [QUOTE="Queen, post: 90610, member: 27"] Neural Amp Modeler's second architecture landed on June 2, 2026, and HeadRush shipped support for it two months later in firmware 5.1.0. Most hardware took longer. Some still have not. The interesting part is that the board reads the old architecture too. You get [B][URL='https://goldmidi.com/community/threads/headrush-added-nam-and-tone3000-to-prime-core-and-flex-core.76020/']both NAM architectures running natively on HeadRush hardware[/URL][/B], which is not something every modeler manages. So when you browse TONE3000 on the touchscreen, you are picking between two generations of the same technology, and the choice is less obvious than the version numbers suggest. [HEADING=2]The model can finally remember 132 milliseconds[/HEADING] A1 handled the final stage one sample at a time, mapping each input value straight to an output value. A2 replaced that with a convolutional head that looks across a window of samples before deciding what comes out. More useful is what happened to the receptive field. A1 could see about 4,100 samples of history. A2 sees roughly 6,350, which works out around 132 milliseconds at 48 kHz. That is where the audible difference lives. Power supply sag, the way a tube amp squashes and recovers after a hard chord, the slow bloom on a cranked circuit. None of that happens inside four milliseconds. It unfolds over a tenth of a second, and A1 could not see far enough back to model it properly. The efficiency came from swapping the tanh activation for a leaky ReLU, which freed enough headroom to grow the network without growing the processing bill. Mixed kernel sizes and retuned dilation patterns did the rest, including killing a ringing artifact that A1's uniform spacing produced on some captures. None of this arrived free for hardware owners. A2 is a new architecture rather than a drop-in patch, so every manufacturer had to build support deliberately, and most of the industry is expected to land it before the end of the year. [HEADING=2]A thousand listeners ranked A2 above the commercial rigs[/HEADING] TONE3000 ran a MUSHRA test, which is the methodology the BBC and EBU use for exactly this kind of question. More than 1,000 participants submitted over 100,000 ratings, scoring anonymized models against real recorded gear on a scale of 0 to 100. The field included A2-Full, A2-Lite, A1-Standard, A1-Nano, Neural DSP V2 and V1, IK Multimedia ToneX V2, and Line 6 Proxy. A2-Full took the highest median scores across the comparisons. The result worth sitting with is A2-Lite. It scored alongside A1-Standard while running in a fraction of the processing budget, which is why it can live on a 600 MHz ARM chip that costs three dollars in bulk. The efficiency gain is not marginal either. A2-Full runs 30 to 40 percent leaner than A1-Standard, which works out at three A2 models for the processing cost of two old ones. Retraining the library moved the median error signal ratio from 0.0062 with A1-Standard to 0.0033 with A2-Full. Roughly half the error, from the same source recordings. [HEADING=2]Your old A1 favorites still load, and most have twins[/HEADING] Nothing breaks. HeadRush accepts one A1 file per rig in Standard, Lite, Feather, or Nano size, so the capture you have been gigging for two years loads exactly as it always did. You probably do not need it anymore. TONE3000 retrained its entire A1 catalog to A2, so the capture you love almost certainly has a newer version sitting in the same place. Look for [B]the A2 badge on the tone pack page[/B], or set the architecture filter before you start browsing. The originals are still there behind an A1 Legacy filter if you want to compare, and comparing is worth doing once. Load both, play the same riff through each, and decide for yourself rather than taking a median score on faith. The A1 size tiers are worth understanding before you go back. Standard was the accurate one. Lite simplified the internal structure while keeping most of the dynamic character. Feather traded accuracy for low latency and smooth real-time response. Nano pruned aggressively and gave up transient response and low-end dynamics to fit on cheap hardware. Nano is the tier to replace first. It existed because small chips could not run anything better, and A2-Lite has made that particular compromise unnecessary. One A2 download covers both variants, so there is no second decision to make at the point of download. The file carries Full and Lite together, and the host picks based on what it has spare. The same holds if you capture your own gear, since one training run produces both variants and takes roughly as long on the same machine as A1 always did. [/QUOTE]

Name