MLVC: Multi-platform Learned Video Codec for Real-World Deployment

A team of researchers has introduced MLVC, a neural video codec engineered to deliver consistent performance across consumer hardware from Apple, Intel, and Qualcomm, overcoming a key barrier that has kept such codecs confined to research labs. The codec, detailed in a paper submitted on 26 Jun 2026, is designed to solve a fundamental problem: existing neural video codecs, while more efficient than classical standards, produce non-deterministic results on different hardware platforms, leading to catastrophic decoding failures [1]. MLVC avoids this by explicitly transmitting scale parameters through a hyperprior, which guarantees entropy coding consistency across devices without requiring bit-exact arithmetic [1]. While this method introduces some bitrate overhead, the team recovered most of the lost efficiency through architectural changes, including gated memory and ReGLU activation, a long-term reference recovery mechanism, and domain-specific perceptual training [1]. On the VCD video conferencing benchmark, MLVC achieved a greater than 70% BD-rate improvement, measured by mean opinion score, over hardware HEVC, which the authors describe as the strongest deployable baseline [1]. Its subjective quality was competitive with DCVC-RT, a neural codec that cannot operate across diverse platforms [1]. Both the encoder and decoder run at an average of 100 FPS on commodity neural processing units from Apple, Intel, and Qualcomm [1]. The paper states this is the first neural video codec to combine competitive compression, real-time speed, and cross-platform robustness on consumer devices [1]. The challenge of cross-platform consistency is not unique to video codecs. A separate study on WebGPU dispatch overhead for large language model inference found that per-operation validation costs vary significantly across GPU vendors and backends, with per-dispatch overhead ranging from 24-36 µs on Vulkan to 32-71 µs on Metal [3]. That research, which tested hardware from NVIDIA, AMD, Apple, and Intel, concluded that per-operation overhead dominates performance in dispatch-heavy pipelines, a finding that underscores the difficulty of achieving uniform inference behavior across heterogeneous hardware [3]. Apple's own history of processor transitions illustrates the complexity of maintaining software compatibility across architectures. The company has switched its Mac line's central processors three times: from Motorola 68000 to PowerPC in 1994, from PowerPC to Intel x86 in 2006, and from Intel to its own ARM-based Apple silicon, a process announced in 2020 and completed in 2023 [7]. Each transition required extensive software adaptation, including binary translation layers like Rosetta, to bridge incompatible instruction set architectures [6]. The MLVC codec's approach to hardware robustness addresses a parallel challenge at the application level, aiming to make a single model viable across the fragmented landscape of consumer NPUs without per-platform tuning.