The M5 Ultra Fuses Four Chips Into One, a First for Apple Silicon

Every Ultra chip Apple has shipped so far has been two Max chips bolted together with an interconnect called UltraFusion. The M5 Ultra is four. That's the whole story, and it's a bigger deal than the core counts make it sound.

Apple announced M5 Ultra today inside the new Mac Studio, where it scales to a 36-core CPU, an 80-core GPU, and 512GB of unified memory. Apple calls it the most powerful chip it has ever made. It's also the most structurally unusual.

How you glue four dies together

M5 Ultra uses UltraFusion to connect two dual-die M5 Max chips, forming a quad-die architecture for the first time in an M-series system on a chip. Apple says the new interconnect pushes inter-die bandwidth past 4.4TB/s and increases connection density by more than 6x.

That density figure is the interesting one. The hard part of stitching dies together is not raw bandwidth, it's latency and coherence: making four separate pieces of silicon behave like one processor so software never has to know or care. If the connection is slow or sparse, you end up with something that acts like four computers in a trench coat. Apple's claim is that the four dies behave as a single unified processor. If that holds up under independent testing, it's the most impressive engineering in today's announcement.

What it adds up to

The CPU tops out at 36 cores: 12 super cores and 24 performance cores. Apple claims 1.25x higher single-threaded and 1.3x higher multithreaded performance than M3 Ultra. Those are modest generational gains, and Apple is not hiding it.

The GPU is where the leap sits. Up to 80 cores, with a Neural Accelerator in every one, the first time Neural Accelerators have appeared on an Ultra chip. Apple claims up to 4.5x the peak GPU compute for AI versus M3 Ultra and over 6x versus M1 Ultra, plus up to 40 percent faster graphics from second-generation Dynamic Caching, hardware-accelerated mesh shading, and third-generation ray tracing.

Memory runs to 512GB at 1.2TB/s, half again the bandwidth of M3 Ultra. The Media Engine doubles the video encode and decode blocks of M5 Max and includes four ProRes engines, which is how Apple reaches its claim of 33 simultaneous streams of 8K ProRes 422 at 30fps. A 32-core Neural Engine handles Apple Intelligence work on device.

Why it matters beyond Mac Studio

The quad-die move tells you where Apple thinks its ceiling is. Apple cannot simply build one enormous die. Yields collapse and costs explode. Stitching dies together is how it keeps scaling, and today it demonstrated it can go from two to four without software noticing. Eight is the obvious next question.

For now the practical upshot is narrower: a machine that holds a very large model entirely in memory and feeds it fast enough to be useful. Whether that's worth what Apple is charging depends entirely on how many hours a day you would keep it busy.