SkyRL has moved its reinforcement learning stack to FP8, a reduced-precision format, and says it matches the convergence of BF16 while shaving up to 23% off rollout step time. The gains were achieved on NVIDIA GPUs, according to the company.
Why FP8 in reinforcement learning
Reinforcement learning is compute-hungry. Agents run thousands of rollout steps to gather experience, and each step adds to the training bill. SkyRL's new stack uses FP8 for those steps, a format that stores numbers in half the bits of BF16. That cuts memory pressure and speeds up the math.
The company didn't say which specific GPUs it used, nor how many. It just said the efficiency gains show up on NVIDIA hardware.
Convergence that matches BF16
Lower precision often means a trade-off. Less accurate numbers can hurt model convergence. SkyRL says its FP8 stack doesn't pay that price. The stack matches BF16 convergence, which is a strong claim because BF16 is already the go-to format for many large-scale training runs.
If the claim holds, it means users get the speed of FP8 without the usual accuracy hit. That's the kind of result that makes engineers sit up, because rollout time is often the bottleneck in RL workloads.
What the 23% actually buys
A 23% reduction in rollout step time adds up fast in long training sessions. A model that takes a week to train could finish in under six days, assuming the gain scales. SkyRL didn't disclose whether the reduction applies to the full training loop or just a single step, nor did it share benchmark details beyond the headline number.
The company also hasn't said if the FP8 stack works on GPUs from other vendors, or if it's tied to NVIDIA's specific architecture. Those questions remain open.
SkyRL's move comes as the industry keeps pushing for faster, cheaper training. FP8 has already appeared in other parts of the AI stack, but using it in the RL rollout phase is less common. The company's announcement is brief, but it points to a practical way to trim training time without rethinking the algorithm.
The missing detail is the measurement. SkyRL hasn't explained how the 23% figure was calculated, which benchmarks it used, or how much variation exists across different models. Those specifics matter for anyone trying to reproduce the result.
