[Bug Report] TiledCamera hangs on RTX 5090 Laptop (GB203) but works on RTX 5090 Desktop (GB202)

[KunHsiang]

Description

TiledCamera hangs indefinitely on RTX 5090 Laptop GPU (GB203), while the same code works perfectly on RTX 5090 Desktop GPU (GB202). Replacing TiledCameraCfg with CameraCfg resolves the issue on the Laptop GPU.

Environment Comparison

	Laptop (HANGS)	Server (WORKS)
GPU	RTX 5090 Laptop (GB203, 24GB)	RTX 5090 Desktop (GB202, 32GB)
Driver	590.48.01	570.211.01
CUDA	13.1	12.8
OS	Ubuntu 22.04, Linux 6.8.0-101	Ubuntu 22.04
Isaac Sim	5.1.0	5.1.0
Isaac Lab	2.3.1	2.3.1
Warp	1.11.1	1.11.1

Symptoms

On the Laptop GPU (GB203):

• Process hangs at 100% CPU with no output or progress
• RSS/GPU memory remains constant
• wchan=0 (CPU busy loop, not I/O wait)
• Hangs for 10+ minutes with no recovery

Investigation Results

1. Warp kernel compilation works fine — standalone reshape_tiled_image kernel compiles and executes successfully (258ms compile + run)
2. Warp correctly detects sm_120 — "cuda:0": "NVIDIA GeForce RTX 5090 Laptop GPU" (sm_120, mempool enabled)
3. sm_120.ptx files exist — precompiled binaries are present
4. The hang occurs in the tiled rendering pipeline (omni.replicator layer), not in the warp kernel itself

Steps to Reproduce

1. Use any Isaac Lab environment with TiledCameraCfg sensors on an RTX 5090 Laptop GPU (GB203)
2. Run the simulation — it will hang during the first render call

Example config:

from isaaclab.sensors import TiledCameraCfg

top_camera: TiledCameraCfg = TiledCameraCfg(
    prim_path="{ENV_REGEX_NS}/top_camera",
    offset=TiledCameraCfg.OffsetCfg(pos=(0.0, -0.065, 0.84), rot=(0.5, 0.5, -0.5, -0.5)),
    data_types=["rgb"],
    spawn=sim_utils.PinholeCameraCfg(
        focal_length=2.12, focus_distance=400.0,
        horizontal_aperture=3.6096,
        clipping_range=(0.01, 20.0),
    ),
    width=640, height=480,
)

Workaround

Replacing TiledCameraCfg with CameraCfg resolves the hang:

from isaaclab.sensors import CameraCfg
# Change all TiledCameraCfg → CameraCfg

Note: Camera and TiledCamera produce different pixel distributions (Issue #493), so this workaround may affect policy training/evaluation consistency.

Key Observation

The same GPU model (RTX 5090) with different chips behaves differently:

• GB202 (Desktop): TiledCamera works ✅
• GB203 (Laptop): TiledCamera hangs ❌

This suggests the issue is specific to the GB203 Laptop variant, possibly related to differences in shader core configuration or the tiled rendering pipeline's interaction with the Laptop GPU's Vulkan/RTX renderer.

Related Issues

• Issue #2483 — RTX 5090 unable to run IsaacLab
• Discussion #2869 — RTX 5090 blurry rendering
• Discussion #3612 — Blackwell PhysX fallback to CPU
• Issue #493 — Camera vs TiledCamera different images

[KunHsiang]

Follow-up: Additional testing (2026-03-14)

1. Driver 590.48.01 still hangs

Updated to the latest NVIDIA driver 590.48.01 (CUDA 13.1) on the laptop — TiledCamera still hangs. Process was killed after 120s timeout with no output beyond env creation.

2. Camera consistency test (patched vs native)

We wrote a consistency test script that runs identical simulation (same garment, seed, 20 home-position steps) on both machines and compares observations.

Setup:

• Same code, same garment (Pant_Short_Seen_0), same seed (42), use_fabric=False, device=cpu
• Laptop: TiledCamera monkey-patched to use Camera internals (per-camera render_product)
• Server: Native TiledCamera (render_product_tiled)

Results:

Observation	Laptop (Patched TiledCamera)	Server (Native TiledCamera)
observation.state	✅ Identical	✅ Identical
observation.images.top_rgb	mean=0, max=0 (all black)	mean=204.9, max=234
observation.images.left_rgb	mean=0, max=0 (all black)	mean=217.5, max=234
observation.images.right_rgb	mean=0, max=0 (all black)	mean=196.2, max=239

• Physics simulation is perfectly consistent (state vectors match to 6 decimal places)
• Camera rendering differs completely — the monkey-patch workaround (using Camera's _initialize_impl on TiledCamera) produces empty buffers in short test scripts, though it works during full eval loops (likely due to additional render cycles)

3. Environment details

	Laptop (HANGS)	Server (WORKS)
GPU	RTX 5090 Laptop (GB203, 24GB)	RTX 5090 Desktop (GB202, 32GB)
Driver	590.48.01	570.211.01
CUDA	13.1	12.8
Isaac Sim	5.1.0	5.1.0
Isaac Lab	0.53.1	0.53.1

This confirms the issue is GB203 chip-specific, not driver-version related. Would appreciate any guidance on whether this is a known omni.replicator limitation or if there's a fix planned.

[RandomOakForest]

Thank you for posting this. The team will review it. We will aim to reply soon as GTC is ongoing.

[KunHsiang]

Hi, following up on this as well. We're using both RTX 5090 Laptop (GB203) and Desktop (GB202) for the ICRA 2026 LeHome Challenge. The TiledCamera issue affects both machines differently, making development difficult.

Any timeline for a fix? We're using standard Camera as workaround but would love to use TiledCamera for better performance. Thanks!