GLM-5.3 hits 21.4x speedup on RTX PRO 6000 via Kimi-Linear Decode kernel

teortaxesTex · x · 2026-08-22

KernelBench-Mega results show GLM-5.3 achieves 21.4x speedup over PyTorch baseline on RTX PRO 6000 using Kimi-Linear Decode, improving from GLM-5.2's 11.1x. The optimization uses a single CUDA loadinline launch, persistent 512-thread CTAs, and 30 grid barriers. It keeps Int4 dequantization inside GEMV and absorbs MLA to avoid materializing the fat KV cache. Benchmark comparisons show GLM-5.3 performing strongly against other leading models.

Original post →

More from Infra

Infra channel →