MPS: Metal Performance Shaders

MPS is Apple's library of pre-built GPU kernels (matmul, convolution, image ops): the closest thing Metal has to cuBLAS/cuDNN, and the baseline your custom kernel has to beat.

CUDA equivalent: cuBLAS/cuDNN in role, not in stature. On NVIDIA, beating cuBLAS at dense GEMM is a research result; the library embodies decades of tuning plus hardware co-design. MPS is good, but the ML ecosystem on Apple Silicon has repeatedly matched or beaten it with open kernels:

• The hand-written m5-gemm kernel hits 13.5 TFLOPS fp32 at 4096² on an M5 Max vs MPS's 11.7, though MPS takes the lead back at 8192² where everything is bandwidth bound.
• llama.cpp deleted its MPS path in 2023 (PR #2615↗) in favor of its own simdgroup kernels: ~88% ALU utilization on a 4096² matmul, with the honest caveat that end-to-end gains were ~40% because non-matmul ops dominate.
• MLX uses its own steel library, not MPS, for core ops.

Why beatable: MPS kernels are general and opaque. No function-constant specialization for your shapes, no fusion with your surrounding ops, no quantized paths for LLM weight formats. The places frameworks still lean on MPS are where its coverage is genuinely hard to replicate (convolution variants, some reductions, small-matrix cases where AMX-backed CPU paths and MPS-tuned dispatch beat naive GPU kernels).

For the kernel reader, MPS's main role is the benchmark floor: MPSMatrixMultiplication is a three-object setup (matrix descriptors + the op) and gives honest GPU-time numbers via command-buffer timestamps; the m5-gemm harness↗ is a compact PyObjC example. If your custom kernel doesn't beat MPS at your shapes, ship MPS. That's not defeat; that's the three questions working as intended.