mccl

mccl · /ˈmɪk.əl/ · Mac Collective Communication Library

Topology-adaptive collective communication for clusters of Apple-Silicon Macs, over unified memory and Thunderbolt. mccl sits between your training code and the hardware: you call a collective, and mccl works out the cluster for you — which links are physically up, how to route over them, and which algorithm is fastest for the size — then runs the reduction on the GPU over unified memory with no host/device copies.

Much of inspiration came from nccl

UPDATE: With wwdc26 came fp8 and fp4 support for m5, will be supporting that soon!

Build

Requirements: macOS on Apple Silicon (Metal), a C++17 clang++, and CMake ≥ 3.20. The Macs are wired by Thunderbolt; the build itself has no third-party dependencies.

$ git clone https://github.com/Lazarus-931/mccl.git
$ cd mccl
$ cmake -S . -B build
$ cmake --build build -j

This produces build/libmccl.a. reduce.metal is compiled at runtime, so there is no offline kernel-build step. Launch one rank per Mac with MCCL_RANK, MCCL_WORLD_SIZE, and a shared MCCL_BOOTSTRAP_IP / MCCL_BOOTSTRAP_PORT — the only required configuration; the topology and routing are discovered.

Install

$ cmake --install build --prefix /usr/local

Installs libmccl.a → <prefix>/lib, the headers → <prefix>/include/mccl, and reduce.metal → <prefix>/share/mccl. A consuming CMake project finds it with find_package(MCCL); or compile by hand:

$ clang++ -std=c++17 app.cc -I<prefix>/include/mccl/include \
      <prefix>/lib/libmccl.a -framework Metal -framework Foundation -o app

If reduce.metal is not beside the binary, point MCCL_METAL_DIR=<prefix>/share/mccl at the installed copy. Tests and benchmarks live in a separate mccl-tests repository.

Performance

batch	64 MB	256 MB	1 GB	4 GB	32 GB
speedup vs 1 Mac	1.0×	1.7×	2.1×	2.7×	2.9×

Structure

src/graph/     topology: discover (interfaces + liveness) -> topo (system + links) -> paths
               (all-pairs widest path) -> search (per algorithm) -> connect (per-rank ring/tree)
               -> tuning (cost model); trees = double-binary builder
src/device/    per-Mac compute: metal (MTLDevice + pipeline cache + UMA reduce), reduce_kernel,
               primitives (send / recv / reduce building blocks), and one file per collective
               (all_reduce, all_gather, reduce_scatter, broadcast, reduce)
src/transport/ m2m: striped TCP over Thunderbolt, plus an RDMA-over-TB backend behind the facade
src/           comm (init -> discover -> graph -> connect; split / finalize / abort), coll (public
               API), bootstrap (rendezvous), socket, allocator (page-aligned UMA)