PerfBench · ARM64 NEON + WGSL

Anthropic-PerfBench

Fused RMSNorm+residual, in-place RoPE, and group-wise INT4 GEMV across three optimization iterations.

3.09×RMSNorm NEON vs scalar
2.35×RoPE NEON vs scalar
11.83×INT4 GEMV NEON vs scalar
66.31 GB/smeasured copy bandwidth
Memory roofline for the three NEON kernels
KernelScalarNEONSpeedup
RMSNorm+residual6.454 ms2.087 ms3.09×
RoPE1.733 ms0.738 ms2.35×
INT4 GEMV20.686 ms1.748 ms11.83×

Apple Silicon NEON results. OCI Ampere perf counters and WebGPU timings not run; roofline caveats are documented in RESULTS.md.

Source & CIPortfolio hub

Measured local evidence only. No deployment, customer, or trained-weight claim is made unless explicitly stated.