2benchesApple silicon
All chips

M3 Ultra

A16 / A17 · M3 · Everest / Sawtooth family · 3 nm · First available 2025-03-12

Every configuration

ConfigurationCPU core layoutBandwidthGB7 CPUGB7 MetalPrefill Q4_0 tok/sDecode Q4_0 tok/s
28 CPU / 60 GPU (binned)20P + 8E819 GB/s35,054225,3731,07388.4
32 CPU / 80 GPU (full chip)24P + 8E819 GB/s39,039249,5641,47192.1

All benchmarks

BenchmarkMedian
Geekbench 7 · Multi-core37,047 (35,054–39,039, 2 sources)
Geekbench 7 · Metal GPU237,469 (225,373–249,564, 2 sources)
Geekbench AI · GPU · Single—
Geekbench AI · GPU · Half—
Geekbench AI · GPU · Quantized—
Geekbench AI · Neural Engine · Single—
Geekbench AI · Neural Engine · Half—
Geekbench AI · Neural Engine · Quantized—
Prefill 512 · F161,330 (1,122–1,538, 2 sources)
Decode 128 · F1641 (39.8–42.2, 2 sources)
Prefill 512 · Q8_01,287 (1,086–1,488, 2 sources)
Decode 128 · Q8_063.7 (63.6–63.9, 2 sources)
Prefill 512 · Q4_01,272 (1,073–1,471, 2 sources)
Decode 128 · Q4_090.3 (88.4–92.1, 2 sources)

Similar performance

Nearest measured configurations by Geekbench 7 CPU and Metal. Ratios above 1 are faster.

28 CPU / 60 GPU (binned)

32 CPU / 80 GPU (full chip)

Machine-readable summary (JSON) · Full dataset