RMSNorm Isn't Necessarily Faster Than LayerNorm

cHHillee · x · 2026-07-13

The author argues that the claim of RMSNorm being significantly "cheaper" than LayerNorm is oversimplified.

The core point is that these operators are often memory-bound kernels; the actual bottleneck is moving data to GPU cores rather than the arithmetic itself, meaning end-to-end latency can be similar. The original paper's insufficient understanding of this memory bottleneck skews its conclusions.

However, the author adds that when pushing performance to the limit—like during matmul epilogue fusion—LayerNorm is harder to fuse due to tracking mean shifts, so it isn't strictly "free." Conclusion: RMSNorm isn't necessarily twice as fast as LayerNorm, but LayerNorm isn't zero-cost either.

Original post →

More from Infra

Infra channel →