# Skild AI open-sources Steel physics solver, says it beats PhysX by up to 13x > Source: > Published: 2026-11-09 23:26:05+00:00 # Skild AI open-sources Steel physics solver, says it beats PhysX by up to 13x **Skild says Steel hit 3.90 million environment steps per second on 20 mixed Isaac Lab tasks at 245,760 environments. At a matched batch size of 61,440, Steel reached 3.61 million steps per second and used about 11 GiB of GPU memory to PhysX's 29 GiB.** By [Ryan Merket](https://runtimewire.com/author/ryan-merket) · Published Primary source: [x.com](https://x.com/SkildAI/status/2119844954893946999) ## Why it matters Steel addresses the cost of simulating many different manipulation tasks at once, a bottleneck for training general-purpose robot policies. Skild has published the code, the solver design, and benchmark conditions, so outside teams can test the peak throughput figure and memory footprint. Skild AI open-sourced Steel on November 9th, a GPU-accelerated rigid-body physics solver built with [NVIDIA](https://x.com/nvidia) for training robots on diverse manipulation tasks. [Skild AI (@SkildAI)](https://x.com/SkildAI) announced the release [in a thread on X](https://x.com/SkildAI/status/2119844954893946999), saying Steel runs up to 13x faster than PhysX while using roughly 3x less GPU memory in its heterogeneous test. When you purchase through links in our articles, [we may earn a small commission](https://runtimewire.com/affiliate-monetization-standards). This doesn’t affect our editorial independence.