This 64-bit Windows build accelerates selected psychoacoustic and bitrate-decision work with NVIDIA CUDA. LAME retains the final MP3 encoding, cVBRb bitrate controls and bitstream generation.
It is intended for compatible NVIDIA systems and advanced users. For the broadest compatibility, use one of the standard CPU builds on the main cVBRb page.
The package contains native CUDA code for the compute capabilities below. Support refers to the compiled target list, not every NVIDIA model released between them.
CUDA target
Architecture
Representative hardware
sm_61
Pascal
GeForce GTX 1050–1080 Ti, TITAN X/Xp, compatible Quadro P cards
GeForce RTX 30 series and RTX A-series workstation cards
sm_89
Ada
GeForce RTX 40 series, RTX Ada professional cards, L4/L40/L40S
sm_120
Blackwell
GeForce RTX 50 series and compatible RTX PRO Blackwell cards
Not every CUDA-capable NVIDIA GPU is included. Volta, Hopper, Jetson-specific and other unlisted compute capabilities should not be assumed to work with this package.
What the specialist build provides
CUDA-assisted encoding
Selected psychoacoustic masking and bitrate-decision work is offloaded while the CPU retains the sequential LAME encoder and final frame generation.
cVBRb controls
Supports strict VBR minimums, optional maximum bitrate and --bitrate-boost levels for the maintained cVBRb branch.
Conventional CLI output
Provides LAME-style progress, bitrate distribution, LR/MS statistics and block-type statistics, with engineering counters available through --cuda-diagnostics.
CPU-based MP3 decode
--decode uses mpg123 and bypasses CUDA initialization, so decoding does not require an NVIDIA GPU.
Command examples
The executable is named lame.exe in the public package.
EAC supplies its own temporary input and output filenames when the LAME MP3 Encoder parameter scheme is selected.
Decode MP3 to WAV
lame.exe --decode input.mp3 output.wav
Decode mode is CPU-based and does not initialize CUDA.
Engineering diagnostics
lame.exe --cuda-diagnostics ...
Add the switch to expose CUDA validation, Graph and authority counters.
Important limitations
Specialist, not default
The standard CPU packages remain the recommended choice for general use, mixed hardware and redistribution to unknown systems.
Hardware-dependent performance
Results vary with GPU generation, driver, CPU, PCIe topology, Windows scheduling and power-management settings.
Mapped-memory requirement
The active CUDA authority mode requires mapped pinned host-memory support. The encoder reports a clear error if the selected environment cannot provide it.
No CPU fallback package
This archive does not include a separate CPU-only encoder. Download a standard x64 build from the main page when CUDA is unavailable.