One Single System Image across all your Macs

Turn the Macs you own
into one computer.

mac-ssi fuses the CPU, GPU, Neural Engine, memory and storage of every Mac you own into a single compute fabric — over Thunderbolt 5, Ethernet, or Wi-Fi. Run any workload across all of them, unchanged.

mac-ssi compute fabric
608 GBunified RAM, 4 Macs
1 PIDone process, every node
TB5·LAN·Wi-Fiany link works
10–100×lower cross-node latency
The whole point

Run anything across the pool.

You own more Macs than you use at once. mac-ssi makes them act as one machine — memory spans every node's RAM, compute lands on whichever Mac has idle cores, GPU or ANE, a job appears as a single PID. Unmodified software just sees a bigger computer.

# run any binary on the best node — by free CPU/GPU/RAM, locality, or energy
$ ssi run ./render_batch --gpu --min-mem 65536

# place work where it costs the least power
$ ssi run --mode energy ./nightly_job

# one process table across every Mac
$ ssi ps

Rendering, simulation, builds, data processing, batch compute, training, inference — anything bottlenecked by a single machine gets the whole pool.

90-second tour

See it in action.

One machine, made of many

A real single system image.

🧠

Aggregate everything

CPU, GPU, Neural Engine, RAM and storage pooled across nodes. One PID, one memory space, one filesystem — no distributed-programming model.

🔗

Any link

Thunderbolt 5 with RDMA for max throughput, plain Ethernet, or just shared Wi-Fi. The fabric adapts to whatever connects your Macs.

🧩

Distributed shared memory

Allocate across the pool's combined RAM with MOESI coherence and per-region consistency. Load data bigger than any one Mac.

📍

Locality-aware scheduling

Predictive prefetch, write coalescing, tiered hot/cold memory, and run-where-the-data-lives placement cut cross-node latency 10–100×.

🌱

Energy-aware

Joules metered per node. --mode energy runs work where it costs the least power — Apple Silicon's perf/watt, pooled.

🚀

Unmodified software

No MPI, no job manager, no rewrite. ssi run ./anything is ssh for your whole pool, with a scheduler.

How it works

Closing the interconnect gap.

Making separate machines act as one is bound by the link between them. Apple's on-die UltraFusion runs at 2.5 TB/s; the wire between Macs is far slower — 10 GB/s on Thunderbolt 5, less on Ethernet or Wi-Fi. mac-ssi closes that gap in software so the pool behaves like one computer.

Predictive page prefetch Write coalescing Tiered hot/cold memory Locality scheduling RDMA buffer pooling LZ4 page compression MOESI coherence
single compute fabric
One example workload

Large-model inference, for free.

Because the fabric pools GPU memory, a model too big for any single Mac just runs across the cluster — sharded automatically, served behind one OpenAI-compatible endpoint. It's an application of mac-ssi, not the point of it.

$ ssi serve some-large-model --port 8080
$ curl localhost:8080/v1/chat/completions -d '{"messages":[...]}'   # plain OpenAI API
API & CLI

Everything is one ssi command.

CommandWhat it does
ssi run ./workloadRun any binary on the best node, scheduled by CPU / GPU / ANE / RAM / locality / energy
ssi ps · killOne process table across every Mac
ssi memoryDistributed shared memory across the pool's combined RAM
ssi gpu · anePool and dispatch to GPU / Neural Engine across nodes
ssi fsOne virtual filesystem spanning every node's storage
ssi status · nodes · resources · topologySee the pool — and the fabric links — as one machine
ssi serve <model>example workload — distributed inference → one OpenAI endpoint
Get started

Install on every Mac.

$ brew install --cask openie-dev/mac-ssi/mac-ssi
$ ssi up       # start the agent — peers auto-discover
$ ssi status   # your pool, as one machine

Apple Silicon · connects over Thunderbolt 5 (RDMA), Ethernet, or Wi-Fi · macOS 26.2+ unlocks TB5 RDMA.