The new GPU collision detection method guarantees no penetrations throughout integration for VBD-like solvers (although you can also use it with Newton-based solvers with less optimal results). Over the past few years the direction for graphics research have mainly been looking at accelerating the IPC solver (which theoretically guaranteed penetration-free simulation but was too slow to use for real-time simulation), but seems researchers have figured out an alternative path that doesn't have the same scale-up issues on the GPU. Although I think you would still need an IPC-like second-order solver for very accurate simulations (ex. for mechanical engineering), for games and VFX this is certainly good enough.