Thirty 2012-era Apple Mac Minis, networked into a single parallel cluster, delivered 1.3 teraflops in measured benchmarks at Spain's Museo de Historia de la Computación. The 18-month build produced a 1:1-scale visual replica of the Cray-1, the vector-processing supercomputer Cray Research released in 1976.

The museum's array executes its workload across 64 active cores split between five quad-core Intel Core i7 units and 25 dual-core Core i5 machines. Twenty-eight of the 30 Mac Minis carry 2012 date codes; the remaining two ship from 2014.

Developers measured the 1.3-tflops figure with a custom matrix-multiplication Python script. An earlier configuration with all 70 cores online had registered 1.5 teraflops before one node went offline.

How does the replica compare to the 1976 original?

The Cray-1 applied a single instruction across multiple data elements simultaneously through dedicated vector registers. Its peak theoretical performance sat at approximately 160 megaflops.

The Mac Mini array takes the opposite architectural route: independent processors share work through a distributed cluster, the execution model that now dominates hyperscale cloud infrastructure. That throughput gap, calculated from the museum's measurement against the 1976 theoretical ceiling, works out to roughly 8,000× between the 1.3-tflops result and the original machine's 160-megaflop peak. The figure underlines a five-decade shift from specialized single-thread pipelines to commodity multi-core throughput.

The original machine executed vector instructions sequentially through dedicated registers optimized for tight loops over contiguous data. The replica breaks identical workloads into independent units dispatched across discrete CPUs connected by Ethernet. That architectural inversion — single fast pipeline versus many slow parallel ones — defines how high-performance computing has been built for the last two decades.

What does the project actually demonstrate?

The replica marks the 50th anniversary of both the Cray-1 launch and Apple's founding year. Museo staff positioned the build as a teaching tool that lets visitors compare purpose-built vector hardware with commodity parallel hardware inside a single working exhibit.

The benchmark choice — a matrix-multiplication Python script written in-house — was deliberate. Matrix workloads map cleanly onto both vector and parallel execution models, giving visitors a side-by-side comparison of two architectural philosophies that dominated different eras of high-performance computing.

How was the cluster assembled?

The developers worked with what they could source. The mix of five quad-core i7s and 25 dual-core i5s creates heterogeneous core counts across the array, and the cluster software must schedule around them.

Losing one node dropped the active core count from 70 to 64 and cost roughly 13% of measured throughput — a reminder that fault tolerance and workload balancing now matter as much as peak silicon performance in parallel systems. The two 2014 units sit inside a fleet that is otherwise 2012 vintage. That pattern mirrors how real-world HPC clusters evolve: hardware fleets age unevenly, and operators keep arrays running by mixing generations.

What's next for the museum?

The exhibit stands as a working display at the Museo de Historia de la Computación. No commercial deployment, hardware expansion, or replication program has been announced.

As the museum's measured throughput moves back toward the 1.5-teraflops mark once the offline node returns, the project will continue to illustrate how the cost-performance curve of compute has reset five decades after Cray-1's release.