NVIDIA puts Isaac Sim and L40S GPUs on AWS to double robot simulation speeds

Training robots in the physical world is painfully slow, expensive, and risky if hardware breaks.

NVIDIA just brought its Isaac Sim robotics platform and L40S GPUs to Amazon Web Services, letting developers build, test, and deploy AI robots directly in the cloud.

Doubling simulation speeds on Ada Lovelace

The update centers on the NVIDIA L40S GPU, now accessible through preconfigured Amazon Machine Images on AWS Marketplace.

Built on the Ada Lovelace architecture, the L40S delivers up to a 3.8x rendering leap for Omniverse compared to the previous generation. Across broad robotic simulation workloads inside Isaac Sim, that translates to a 2x performance jump over the older A40.

Beyond physics and rendering, the chip handles generative AI workloads. NVIDIA notes it can fine-tune large language models in a few hours and power real-time inference for chat and text-to-image tools.

Isaac Sim also includes Omniverse Replicator, a synthetic data generation engine that produces training data for robotic perception models, bypassing tedious real-world data collection.

Amazon builds virtual twins for warehouse fleets

Demand for warehouse automation is accelerating fast. ABI Research projects worldwide mobile robot revenue in warehouses will surge from $11.6 billion in 2023 to $42.2 billion by 2030.

Amazon Robotics, which has deployed over 750,000 robots across its global facilities, is among the first to adopt the setup on AWS.

The retail giant is using Omniverse to create digital twins of its fulfillment centers, planning and testing layouts virtually before physical buildout. It is also tapping Isaac Sim for sensor emulation to speed up development of its Proteus autonomous mobile robot.

Other early enterprise adopters include Soft Robotics and Theory Studios.

Why it matters

We think putting Isaac Sim and L40S on AWS is a brilliant move to democratize physical AI development.

Running high-fidelity physics and ray-traced sensor simulations historically required specialized local clusters. Moving that stack into on-demand cloud instances means early-stage robotics startups now get access to the same virtual training grounds used by Amazon.

For NVIDIA, this secures its grip on the industrial metaverse. The L40S does not compete with top-tier accelerators like the H100 for core model pretraining; instead, it establishes an essential middle ground for graphics, synthetic data, and real-time inference.

The lingering challenge remains the sim-to-real gap, as edge cases in real warehouses can still confound virtual models. But cutting simulation times in half is a massive step toward bridging that divide.