By Evan Vega

Exo Labs, Apple and independent testers all publish figures on Mac AI clustering. They measure different things.

Claims about Apple silicon AI clustering come from two very different places, and the
distinction is routinely lost. Exo Labs, which makes the clustering software, publishes speedups of
up to 1.8× on two devices and 3.2× on four. Apple reports up to 4.6× faster
time-to-first-token on an M5 Ultra cluster.

Independent testing by Jeff Geerling on a loaned four-machine cluster measured a large open
model at roughly 32 tokens per second. Vendor figures and independent measurements are both useful
and are not interchangeable. Novel Cognition labels each by origin throughout its explainer at
macluster.novcog.us.com.

Read the full investigation →

Related: Frontier Watch

Read the full explainer →