RigSuite 0.96 Release Notes

• Stereo Multi-Core: the two amp models of a stereo rig now run concurrently on two CPU cores, and a stereo pedal-NAM pair splits the same way. This fixes crackling and dropouts at high oversampling in stereo. The sound is identical to the last sample; there is nothing to configure.
• Phase Lanes: one oversampled amp model now splits its work across up to eight CPU cores, sample-for-sample identical to running it whole. This ends the occasional random stutter while live monitoring at 384k: in the verification session, a quarter-million live audio blocks passed without a single missed deadline. On by default, nothing to configure.
• The neural engine got faster: a fast inference kernel for A2-class models and modern CPU vector instructions land together, roughly halving the CPU a full-quality amp costs at 8x oversampling (about twice the speed). In practice that is more amp instances per project, more headroom to run stereo rigs and high oversampling without the meter climbing, and a bigger drop still when you switch an A2 model to Lite or turn on Live Performance. It is also why the Windows build now needs a CPU from 2011 or later.
• Every library learned to be browsed: the CAB page gains arrow-key audition (step through cabs and hear each one), sorting, and a rescan that never loses your place; the pedal browser gains search, favorites, recents, and in-place folder management; and the convolution reverb gets its first real browser, showing each IR's measured character, decay time, and true-stereo status from the Workshop's analysis. Presets round it out: the preset menu can now rename, duplicate, and delete, and duplicating a factory preset gives you your own editable copy.
• The Reverie freeze becomes a real drone instrument: a new BODY control sits next to the 8ve/5th chips. Turned up, a frozen chord keeps its body and its notes instead of slowly dissolving (measured, the lowest note of a held chord now loses about 1 dB in a minute where it used to lose everything), and the shimmer sings as a stable octave choir above the chord instead of washing it out. Freeze rides a MIDI switch ("Delay freeze (hold / re-grab)" in the MIDI menu), and it learned to layer: tap to freeze, tap again while frozen and the held drone melts into a new capture of whatever you are playing now, over about two seconds (a quick double tap releases; a latching pedal still freezes only while held). With BODY on the expression-pedal list too, the whole drone instrument works from a pedalboard.
• A blank preset and a per-pedal reset: the preset menu gains "New blank preset" (a clean mono slate: nothing loaded, every knob at its default, your calibration untouched), and every FX pedal gains a small RESET button that returns that one pedal's knobs to their defaults while keeping your engine/mode choice and its on/off state. Settings still deliberately stick when you switch a pedal's modes; RESET is the quick way back.
• The oversampling slider tells the truth: the four-way switch is now a detented slider labeled with the actual rendering rate for your session and model (96k, 192k, 384k at a 48 kHz session), and a model that can only run at its native rate says so instead of offering choices that do nothing.
• Anti-alias, in plain terms: the four filter options are now three, Auto / Low latency / Linear phase, because measurement showed the real differences are latency and phase, not tone. Auto is the new default: zero latency live, linear phase for renders.
• Renders can go to 32x: a new Render OS setting raises Render Mode MAXIMUM's oversampling ceiling to 16x or 32x. It applies only to offline renders, so live playback and your presets are untouched.
• The pedal NAM gets its architecture control: the same MODEL ARCHITECTURE chip as the amp tab, on the DRIVE panel. A2 pedals switch between Full and Lite; every pedal shows its class.
• A timing fix for 44.1 kHz sessions: the delay the plugin reports to your DAW is now sample-exact at every oversampling setting. On 44.1 and 88.2 kHz sessions it could be off by roughly a hundred samples with some filter choices, enough to smear a doubled track.
• "Check for updates" works again: the manual check button in Settings had become unreachable through a layout bug (the automatic daily check was unaffected). It now sits on the Support row, and the version readout joined the copyright line.
• macOS fixes from August, folded in: a hardening pass that fixed a Logic Pro crash on model and IR load, and the Workshop's macOS dialog buttons that answer again (which restores the "click to fix" embed prompt a Mac user was stuck on). Both reach every platform in 0.96.

Stereo Multi-Core: two amps, two cores

Since oversampling arrived, some Windows machines could not run a stereo rig at 8x without the audio crackling or collapsing into noise, even though the DAW's CPU meter read a few percent. The meter was telling the truth and hiding it at the same time. A CPU meter shows an average; audio lives on deadlines. Every buffer, the whole rig has to finish before the sound card needs the next block, and all of RigSculpt's work for one instance ran on one core. Two oversampled amp models on one core can miss that per-block deadline while averaging four percent over a second. That is also why it glitched on silent input: a neural model costs the same to run on zeros as on a riff.

The fix is to stop asking one core to do both amps. Amp B's model now runs on its own worker thread, at pro-audio priority, concurrently with amp A on the audio thread. Under STEREO IN with a pedal NAM on each side, the pedal pair splits across the cores the same way. The two halves are then joined and the block continues exactly as before.

• The sound is identical. Each model computes exactly what it computed before, on the same samples, in the same order. Running them side by side changes when the work finishes, never the result. The output is bit-for-bit the same.
• It is always on. There is no toggle and no thread count, because with two models there is exactly one useful way to split them. The handoff costs microseconds; what it buys is the deadline headroom that was the whole problem.
• Hybrid CPUs are handled. On Windows the worker registers with the pro-audio scheduler and explicitly declines efficiency-core parking, so recent Intel machines keep both models on performance cores.
• Mono rigs are untouched. One model has nothing to split; the worker sits idle at zero cost until a second model exists.

Credit where it is due: the decisive evidence came from DLC86's NAM-Oversampler, whose multi-core feature cleared the same wall on the same machine and proved the collapse was a single-core throughput cliff, not a defect. Stereo Multi-Core is RigSculpt's own design and code, and it splits along a different seam than his: whole models in parallel, which fits a plug-in that already has two or three of them. (His seam, splitting inside one model, arrives in this release too, as Phase Lanes below.) The Windows thread-configuration recipe was informed by his implementation, and the third-party notices say all of this precisely.

One honest note: your DAW's per-track CPU readout may show a slightly higher number than before, because two cores are now reporting work instead of one queuing it. The machine is doing the same math with room to breathe.

A faster neural engine

Stereo Multi-Core put each amp model on its own core, but one model at 8x oversampling is still a big job for a single core, and on some machines its worst-case block could still brush the audio deadline: an occasional stutter at 384k even with the average CPU reading low. Two build-level changes narrow that gap directly.

• The fast A2 kernel. A2-class models have a fixed, known network shape, and a kernel that exploits that (fixed-size matrix tiles instead of general-purpose ones) computes the same mathematics faster. The kernel comes from DLC86's NAM-Oversampler core and is credited in the third-party notices; RigSculpt's copy carries its own oversampling support. It had been sitting in RigSculpt's engine, switched off; 0.96 switches it on and opens its model gate so real captures actually use it.
• AVX vector instructions. The Windows build now compiles with AVX, letting the math library process eight numbers at a stroke instead of four, across every model architecture, including classic WaveNet captures the fast kernel does not serve.

Measured together on the machine that stuttered: about 1.9x faster on a full-quality model at 8x. Lite-architecture models gain even more, about 2.6x, which has a practical consequence worth knowing: switching an A2 model to Lite, or engaging Live Performance mode, now buys a much larger CPU drop than it used to. The toggle finally earns its keep. The output is equivalent to the old engine to well below audibility: the difference measures at −126 dB, far under the level of the anti-alias filters themselves. Raw speed alone did not fully cure the live stutter, though, and profiling inside the DAW showed why; that story, and its fix, is the next card.

System requirement: because of AVX, the Windows build now requires an AVX-capable CPU: any mainstream Intel or AMD processor from 2011 on. The installer checks for AVX and says so plainly, rather than letting an unsupported machine find out the hard way. The macOS builds are unaffected.

Phase Lanes: one amp, eight cores

Profiling a live 384k session on the machine that stuttered showed something the isolated benchmarks missed: inside a DAW, the model ran well over twice as slow as the same model on a test bench. The reason is memory, not arithmetic. An oversampled model's state is eight times larger than the model's native state, and streaming that through a CPU cache every two milliseconds, between everything else a DAW does, is what actually set the pace. Making the math faster helped; it could not change how much memory one core had to touch.

Phase Lanes changes how much memory one core has to touch. At 8x oversampling, the model's temporal structure makes the oversampled stream separate exactly into eight interleaved native-rate streams. Each lane runs its own copy of the model at native rate, on its own core, with a working set an eighth of the size, which fits in a core's cache. The lanes' outputs interleave back into the oversampled stream.

• The split is exact. This is not an approximation: the decomposition is a property of how these networks handle time, and the lane outputs are sample-for-sample identical to running the model whole. Our test harness proves the two paths bit-identical across block sizes, lane counts, and mid-stream architecture switches.
• Verified where it hurt. On the machine this round was about, the same evening, same rig: lanes off, hundreds of missed audio deadlines in a few minutes of live playing; lanes on, a quarter-million live blocks and not one miss.
• On by default, self-managing. Lanes engage where they apply (an oversampled, non-LSTM model on a mono path; stereo rigs lane each side) and stand down everywhere else. If the session's sample rate changes, the plug-in notices the split no longer fits and rebuilds it for the new rate on its own.
• Model loading stays quick. Arming the lanes means building one model copy per lane. The file is read and decoded once and the copies are built on parallel cores, measured about four times faster than the naive approach; loading an amp at 8x takes slightly longer than 0.95, not multiples of it.

Credit, precisely: the insight that an oversampled model separates into independent phase lanes comes from DLC86's NAM-Oversampler, and it is his. RigSculpt's implementation is written from scratch and differs where the two projects differ; the third-party notices state the split exactly.

The oversampling slider, labeled with real rates

A neural model can only run at whole-number multiples of the rate it was trained at. That is a fact about the models, not a limitation of the plug-in, and it means a fractional oversampling knob would be decoration: there is nothing between 2x and 4x for the engine to do. So instead of pretending, the new slider offers the settings that exist and names what each one actually does on your session.

• The stops show rendering rates, not multipliers. At a 48 kHz session with a 48 kHz model: OFF, 96k, 192k, 384k. On a 44.1 kHz session the achievable rates differ, and the labels change with them. What you read is the rate the model really runs at.
• A native-only model says so. LSTM-family models run correctly only at their trained rate. Loading one dims the slider and shows "native rate only" instead of four choices that would silently do nothing.
• Nothing moved underneath. The setting itself, and what each position does, is unchanged from 0.95. Presets and automation are unaffected; this is the same control telling you more.

Anti-alias: three honest choices

The anti-alias section offered four filter options, and choosing between them was murky because the differences were smaller than the names implied. Measurement settled it: across a full grid of round-trip null tests, every option measures tonally identical, nulling at −96 dB RMS or better, and two of the four produce literally the same filter. The real, hearable-in-use differences are two: how much latency the filter adds, and whether it is phase-linear.

The control now says exactly that:

• Auto (the new default): zero-latency minimum phase while you play, linear phase when you render. The right answer for both contexts, chosen by context.
• Low latency: minimum phase, always. For tracking, and for hosts where plugin delay matters.
• Linear phase: phase-linear filtering, always, at the cost of a small fixed latency.

Settings saved by older builds map onto their family, so nothing you chose is lost. And the filters themselves were re-tuned from the same measurement grid: the passband edge at a 48 kHz session moves from about 21.4 kHz to about 22.8 kHz, so less top end is shaved ahead of the model. The retune was chosen by measurement, and every configuration still passes the same null-test bar.

Render oversampling: 8x, 16x, 32x

Live oversampling in RigSculpt tops out at 8x, deliberately: live playback has a deadline, and the envelope that works on real machines stops there. An offline render has no deadline. The new Render OS setting lets Render Mode MAXIMUM take a bounce to 16x or 32x, and the Render Mode caption states the ceiling it will use.

The setting is a preference, not a preset parameter. Presets, host automation, and live playback are all untouched by it; it exists only inside the render. Live modest, render maximal.

Whether 16x or 32x audibly improves a given model over 8x is model-dependent and often subtle. It costs only render time to find out.

The pedal NAM's architecture chip

The amp tab has had a MODEL ARCHITECTURE chip since A2 models arrived: A2 Full or A2 Lite per preset, with the governed state shown in amber. The pedal NAM now carries the same chip on the DRIVE panel, beside the calibration badge.

• A2 pedals switch Full/Lite, saved per preset, exactly like the amps.
• Every pedal shows its class, so what is running is never a mystery.
• Two quiet behaviors end. The pedal used to inherit amp A's architecture choice without saying so, and a live Performance/Quality flip never re-applied to a loaded pedal until it was reloaded. Both now behave exactly like the amps.

Sample-exact latency reporting

Your DAW lines tracks up by asking each plug-in how much delay it adds. RigSculpt's answer is now derived from the actual filters in the path, at every oversampling factor and every filter choice. Before this build, on 44.1 and 88.2 kHz sessions, one filter choice reported no delay at all and another reported a fixed approximation that was only accurate at 2x; either way, a bounce could sit roughly a hundred samples out of time against tightly aligned tracks. Small enough to miss on a solo track, big enough to smear a double.

Settings, repaired and tidied

• "Check for updates" is back. The manual check button and the build-info line at the bottom of Settings had been unreachable since 0.93: a layout bug drew them where no click could land. The automatic daily check was never affected, which is why update notices kept working. The button now sits on the Support row where it belongs, and the version readout is part of the copyright line at the bottom of the page.
• The anti-alias chips fit their labels. The three new choices are spaced by the width of their text, so "Linear phase" no longer runs off the edge.
• Render Mode says what MAXIMUM means. The caption tracks the Render OS setting, so the mode's real ceiling is stated where you choose it.