Apple fused two M1 Max chips into one M1 Ultra with a silicon bridge called UltraFusion

Apple's M1 Ultra is a Frankenstein's monster in a good way: two identical M1 Max chips bolted together using UltraFusion, a silicon interface that serves as a bridge, WIRED's Will Knight reported in April 2022. For practical purposes it acts like a single, impossibly large slice of silicon. 'This technology showed up at just the right time... In a sense, it is about Moore's law,' Tim Millet, Apple's vice president of hardware technologies, told WIRED.

The design was hidden in plain sight. When the M1 Max launched in October 2021, onlookers noticed a stretch of mystery silicon along one edge that appeared to do nothing — it turned out to be one half of the UltraFusion interconnect. TSMC, which manufactures Apple's designs, had begun perfecting die-joining, mostly for cores with different jobs; Apple customized the technique to connect entire chips, what WIRED called the radical version of the chiplet approach.

UltraFusion arrived as the industry pivoted to modular silicon: Intel pushed stacked chiplets, AMD already used TSMC 3D stacking, and Intel, AMD, Samsung, TSMC and ARM announced a chiplet-standard consortium the month before. 'Progress in the future is going to come not only from Moore's law but from creating systems that could be fabricated by different technologies yet to be brought together,' MIT professor Jesús del Alamo told WIRED.

Transistors were getting harder to shrink and making an individual chip much bigger had become impractical.

Two proven M1 Max dies were already in production, so scaling meant assembly rather than a bet-the-company new design.

Embedding half the interface in every M1 Max turned seemingly idle silicon into an option on a future flagship product.

The fused pair presents itself to software as one machine with shared memory, not a cluster of two chips.

When one axis of progress stalls, buy progress on another: compose systems from proven pieces so scaling becomes an assembly problem instead of a riskier redesign from scratch.

The M1 Ultra shipped in the Mac Studio, and benchmarking reported by WIRED put it competitive with the fastest high-end computer chips and graphics processors on the market. Modular designs are changing chip economics: as University of Michigan engineer Todd Austin noted, extra performance may now come as a premium tier — base chip $300, double chip $600 — rather than free annual gains.