A jump from 483 to 1,485 prompt-processing tokens per second on a 284-billion-parameter model immediately gives the M5 Ultra more substance than a routine generational refresh. The presentation is built around testing whether Apple's performance claims survive contact with real workloads, and the results generally support substantial gains. Just as importantly, the testing reveals where those gains are dramatic, where they are merely solid, and what they cost in money, power, heat, and noise.
The comparison with the M3 Ultra is unusually useful because both machines run the same builds and model files, while the configured systems also use matching 8TB storage capacities. Developer workloads show significant improvement even before local AI becomes the focus: a large .NET compilation falls from 71.7 seconds to 52.4 seconds, while the Python Mandelbrot test drops from 10.25 to 5.8 seconds. The presenter sensibly calls the machine overkill for ordinary development rather than treating every benchmark victory as justification for buying a $14,299 configuration.
Storage performance provides another striking result. Sequential read performance rises from roughly 6.5GB/s on the M3 Ultra to nearly 14.9GB/s on the M5 Ultra, while write performance is reported as even more dramatically improved after a questionable initial result is rerun. Correcting that result openly strengthens the presentation, although a fuller explanation of the anomalous first measurement would have made the storage comparison more rigorous. The practical connection to loading enormous local models also keeps the benchmark from becoming an isolated number.
Local AI is where the testing becomes most informative. The explanation of prompt processing as primarily compute-dependent and token generation as heavily influenced by memory bandwidth gives viewers a useful framework for understanding why the new hardware behaves differently. The M5 generation's neural accelerators are then connected to actual software support in llama.cpp rather than simply being treated as another specification-sheet feature. GPU-side memory bandwidth measurements also provide context for the roughly 1.5-times token-generation improvements repeatedly observed across several models.
The strongest section tests Apple's claim of up to four-times-faster prompt processing rather than stopping once a favorable result appears. A dense Qwen model moves from about 430 to 1,800 prompt-processing tokens per second, while a 14,000-token coding-style prompt takes eight seconds instead of 33. Crucially, the presenter then shows that the scale depends heavily on prompt length: roughly 350 tokens produces only a 1.6-times improvement, around 1,700 reaches 3.4 times, and approximately 4,500 tokens are needed to reach the claimed four-times territory. That distinction turns a headline benchmark into genuinely useful purchasing information for people deciding between casual local chatting and large coding-agent workloads.
The less flattering measurements prevent the review from becoming an extended celebration. Reported generation power rises from about 36 watts to 45 watts with only a modest tokens-per-watt improvement, while an intense GPU workload pushes the M5 Ultra above 400 watts according to the displayed measurements, compared with nearly 200 watts for the M3 Ultra. The newer machine also becomes noticeably more audible and warmer. These are important tradeoffs for workstation buyers, though the power discussion would benefit from more controlled testing because the figures are chip-reported rather than wall-power measurements and the extreme GPU workload is only briefly characterized.
There are still limitations appropriate to a first look. Model behavior varies considerably, some tests use different workload characteristics, and the promised deeper comparison of MLX, llama.cpp, 8-bit models, clustering, and larger memory configurations remains for later coverage. The sponsorship also arrives directly in the middle of the technical progression and temporarily breaks the momentum. Even so, the combination of repeatable comparisons, multiple model types, practical prompt lengths, corrected benchmark data, and explicit attention to disappointing results makes this a considerably more informative examination than simply repeating Apple's advertised multipliers.
Pros
- Tests Apple's performance claims against matching M3 Ultra hardware using the same software builds and model files.
- Connects CPU, storage, memory bandwidth, neural accelerators, prompt processing, and token generation to practical workloads.
- Shows that roughly four-times-faster prompt processing is achievable while clearly demonstrating how heavily the gain depends on prompt length.
- Uses multiple large AI models, including mixture-of-experts and dense architectures, rather than relying on a single favorable benchmark.
- Includes meaningful drawbacks involving price, power consumption, heat, noise, and relatively modest efficiency improvement.
- Openly corrects a suspicious storage benchmark instead of leaving the questionable result unaddressed.
Cons
- The extreme GPU power comparison receives less methodological explanation than the core AI benchmarks and does not measure total wall power.
- Several important areas, including deeper MLX comparisons, 8-bit models, clustering, and 512GB configurations, are deferred to future testing.
- The sponsorship interrupts the technical progression just as the discussion moves into memory bandwidth.
- The $14,299 configuration makes the results highly specialized, and the presentation has limited testing of workloads outside development and local AI.
Careful benchmarking makes a convincing case that the M5 Ultra's biggest advantage is not simply faster generation but dramatically improved processing of large prompts for demanding local-AI workflows. The substantial price, increased power draw, heat, and noise keep those gains firmly in workstation territory, but the presentation does an excellent job of showing exactly which users are likely to notice them.

