Isaac Gym Preview 4 incompatible with RTX 5080 (Blackwell SM_120) — libPhysXGpu_64.so missing SM_120 kernels

Omniverse Isaac Sim
2

Hi @korawit.k

Your analysis is spot on – Isaac Gym Preview 4’s libPhysXGpu_64.so only contains CUDA
kernels up to SM_80, with no embedded PTX, so it cannot run on your RTX 5080 (SM_120).
Unfortunately, Isaac Gym Preview 4 is officially deprecated and will not receive updates:

“Note: This is legacy software. Developers may download and continue to use it, but it is
no longer supported.”
– Isaac Gym - Preview Release | NVIDIA Developer

There is no workaround for this – the binary is closed-source, has no PTX fallback, and
NVIDIA will not release an updated version.

The solution is to migrate to Isaac Lab, which
is the official successor to Isaac Gym:

  • Isaac Lab runs on Isaac Sim 5.1.0, whose PhysX library (libPhysXGpu_64.so) includes
    SM_120 kernels – I verified this directly on an Isaac Sim 5.1.0 installation.
  • It supports Python 3.10/3.11 and PyTorch 2.x with current CUDA versions.
  • It provides the same GPU-parallel RL training workflow you’re using in Isaac Gym.
  • RTX 50-series GPUs are fully supported (the “no PhysX” concern about RTX 50 series only
    applies to the old 32-bit PhysX, not the 64-bit version used by Isaac Sim – confirmed
    by NVIDIA    ).

Migration guide for moving from Isaac Gym to Isaac Lab:

For your humanoid locomotion project, Isaac Lab’s omni.isaac.lab_tasks provides equivalent
environments. The migration is straightforward for most RL training code.

To get started:

  1. Install Isaac Sim 5.1.0: Download Isaac Sim — Isaac Sim Documentation
  2. Install Isaac Lab: Local Installation — Isaac Lab Documentation
  3. Follow the migration guide above to port your EngineAI PM01 training setup.