M5 Max

The UltraFusion inter-chip bandwidth of the M5 Ultra has increased from approximately 2.5 TB/s on the M3 Ultra to 4.4 TB/s, with a 6x improvement in connection density.

Under the M5 Ultra's 1.2 TB/s bandwidth, the theoretical speed of a 100 GB quantized model is approximately 12 tokens/s, a 200 GB model approximately 6 tokens/s, and a 300 GB model approximately 4 tokens/s.

If bandwidth reached 2.4 TB/s, the theoretical speed ceiling for a 100 GB model could reach 24 tokens/s, a 60 GB model 40 tokens/s, and a 35 GB model close to 69 tokens/s.

The M5 Ultra's peak AI compute performance is more than 4x higher than that of the M3 Ultra.

For undergraduate-level AI coursework, an M5 MacBook Pro with 24GB memory and 1TB storage is more than sufficient for tasks such as writing code, running small-scale experiments, and data preprocessing.

The M4 Max chip has a memory bandwidth of up to 546 GB/s, more than 4 times the memory bandwidth of mainstream DDR5 system memory.

The M5 Ultra is very likely to adopt UltraFusion dual-chip interconnect packaging technology, combining two M5 Max-level chips into a single logical unit.

LFM2.5-24B-A2B achieves a decoding speed of 184 tokens/s on an Apple M5 Max.

In the time-to-first-token test on a 70-billion-parameter model, the M3 Ultra took 31.6 seconds while the M5 Ultra took only 7.7 seconds, achieving approximately 4.1x, close to Apple's claimed 4.3x.

The actual benchmark result of running a 17GB model on an M5 Max is approximately 31 tokens/second.

The Hacks.gr testers used benchmark programs to compare the Asus ProArt with the 2026 16-inch MacBook Pro (M5 Max processor) and the HP ZBook Ultra G1a 14, but did not disclose specific benchmark numbers.

Nimble 9B completes a decision in an average of 90 milliseconds on the M5 Max, with only 1 to 4 output tokens and between 155 and 841 input tokens.

The M4 Max of the same name comes in two bandwidth variants—410 GB/s and 546 GB/s—and the M5 Max exists in 460 and 414 versions.

The Neural Accelerator requires supporting software to function; running M5 Max with a 2023 old version of llama.cpp yields a read speed of only approximately 990 tokens/s.

Running a 27B model on M1 Max with a 32,000-token prompt requires waiting more than 500 seconds before the first character is output, whereas the M5 Max takes less than one minute (approximately 40 seconds with optimized software).

The decode write speed improvement from M4 Max to M5 Max is only 25%, constrained by memory bandwidth.

In terms of prompt processing speed, the RTX 4090 achieves approximately 3,000 tokens/s and the M5 Max approximately 800 tokens/s, giving NVIDIA a clear advantage in long-document and codebase input scenarios.

On the M5 Max chip, Lily's prefill speed is 1.23× faster than MLX-LM and its decode speed is 1.35× faster (official data).

At a 4K×4K×16K matrix multiplication scale, the M5 Ultra achieved 91 TFLOPS, while the M5 Max and M3 Ultra achieved 34 and 22 TFLOPS respectively.

For small-to-medium-scale models, the M5 Max's powerful compute is redundant and does not offer a cost-performance advantage.