GPU NAM 1.2.4 — A2 Lite default + editor-teardown crash fix

Signed + notarized combined installer (Customize pane: Standalone app + AU / VST3 / CLAP, optional sample models & cabinet IR).

Fixes

Editor-teardown crash (important if you're on an older build). Adding the plugin to a track and then deleting it could crash the host (e.g. Ableton Live) — a use-after-free in the editor's display-link idle pump on teardown. AU v3 (macOS + iOS) now serializes editor teardown on the main thread, and the GPU render path re-checks liveness after the idle pump. Verified unaffected for VST3 / CLAP / Standalone.

A2 "Deluxe Reverb" sounded harsh. Packed SlimmableContainer A2 captures now default to the Lite variant, matching the reference. The larger Full variant is faithful but carries ~5 dB more high-frequency energy — great for a dark capture, harsh on a bright one. Pick Full anytime in Settings; an ⓘ explains the trade-off.

Output level. Output Mode now defaults to Normalized (−18 dBFS) so quiet captures don't invite over-cranking, and is a full Raw / Normalized / Calibrated selector.

Cabinet swaps are click-free (crossfaded), and the Settings panel label/control overlap is fixed.

Verify

shasum -a 256 GpuNam-1.2.4.pkg
# ff9cdfcc8bf02a6d05557dafa2af567e3d0aa39449ec6cd4b02939b3283af22a

Gatekeeper: Notarized Developer ID.

GPU NAM 1.1.0 — open source + all-formats installer

GPU NAM is a Neural Amp Modeler (.nam) player that runs amp captures on the GPU, built on Pulp. This release ships the full open-source repository (Pulp pinned as a submodule) alongside a signed, notarized all-formats installer — the plugin now lives here in its own repo rather than inside the Pulp examples tree.

Install (macOS, Apple Silicon)

Download GpuNam-1.1.0.pkg below and run it. It's signed and notarized — Gatekeeper accepts it with no right-click-to-open dance. The installer's Customize pane lets you pick which formats to install:

• AU → ~/Library/Audio/Plug-Ins/Components/
• VST3 → ~/Library/Audio/Plug-Ins/VST3/
• CLAP → ~/Library/Audio/Plug-Ins/CLAP/
• Standalone app → /Applications
• Sample models & cabinet (optional) → /Library/Application Support/GPU NAM

Then rescan plugins in your DAW. A default capture ships inside the bundle, so the plugin makes sound immediately; load your own .nam from the model slot for real amp tones.

What's in it

• Loads WaveNet (A1/A2), ConvNet, Linear, and LSTM .nam captures, plus RTNeural/Keras (.json) GRU/LSTM — one CPU inference substrate.
• Browse controls: ‹ / › step through the models (or cabinets) in a folder — any architecture is detected automatically as you go — and the clear glyph resets a slot. Cabinet IRs swap click-free while audio runs.
• Opt-in GPU engine for the feedforward WaveNet family: the whole forward pass is fused into one GPU submission per block, validated bit-for-bit against the CPU path. CPU is the honest default for small captures; the GPU engine pulls ahead as captures grow.
• Click-free cabinet-IR swaps, host↔model rate resampling, loudness-metadata Normalize mode, and a faithful native editor.

Verify the download

xcrun stapler validate GpuNam-1.1.0.pkg     # "The validate action worked!"
spctl --assess -vv --type install GpuNam-1.1.0.pkg   # accepted / Notarized Developer ID

Build from source

See the README — Pulp is vendored as a git submodule; CLAP + Standalone need no external SDK.

GPU NAM 1.0.0

GPU NAM 1.0.0 — neural amp captures on the GPU

Load a Neural Amp Modeler .nam capture and run its WaveNet inference to reamp your signal. CPU by default, with an opt-in GPU engine running the exact same model.

Honest about the GPU

Naive per-sample GPU inference is terrible (round-trip per sample). Done right — block-parallel (the network is feedforward, so a whole block computes in parallel) and fused into one GPU submit per block — it wins, and the win grows with model size. The GPU output is validated bit-for-bit against the CPU reference (cross-correlation 1.0). A standard capture runs real-time on either engine (GPU with headroom); a large capture (48-channel, two 16-layer stacks) runs ~9.5× faster on the GPU — the regime where the CPU can't keep up. The status line shows the live per-block cost and how much of real-time it's using.

A small example model is bundled so it runs out of the box; load your own .nam captures for real amp tones.

One notarized installer (Customize → AU / VST3 / CLAP / Standalone). Apple Silicon; run your host natively (not Rosetta).

Plays the open MIT Neural Amp Modeler format; inference is an independent implementation. Bundled example model from sdatkinson/NeuralAmpModelerCore (MIT), with attribution.