Apple M silicon benchmarks

[hifivoice]

On Reddit there is a visualisation of the performance of the M-cores clustered by families, and differentiating the variants by giving them a unique colour. I like this one, as it gives a quick overview of the process node vs variant performance advantages. I copied this idea to my own overviews, and update it with the M5.


Geekbench 6
Geekbench 6 reports separate scores for CPU Single-Core, CPU Multi-Core, and GPU (Metal/OpenCL). Each reflects a different execution model and system bottleneck. Each domain stresses a different aspect of the system and should be interpreted accordingly.

Compute domain	Optimised for	Primary limitation
CPU (Single-Core)	Latency & responsiveness	Instruction throughput per core
CPU (Multi-Core)	Parallel compute scaling	Core count & shared resources
GPU (Metal)	Throughput & bandwidth	Memory system & GPU width

Below you can find the charts with the benchmark results, where the single-core and multi-core are combined in one chart (the M5Pro and Max are estimates).

Geekbench AI
Geekbench AI reports separate scores for CPU, GPU, and NPU, reflecting fundamentally different computing paradigms:

Core-engine	Optimized for	Limitation
CPU	Flexibility & latency	Energy & bandwidth
GPU	Throughput & bandwidth	Efficiency & latency
NPU	Efficiency & specialization	Model/operator coverage

Geekbench AI supports three precision modes. These are not incremental optimisations, but represent fundamentally distinct execution strategies across modern compute engines. In real-world AI inference, FP16 and INT8 are by far the most relevant.

Precision	Why it exists	Primary bottleneck	Dominant engine
FP32 (single)	Maximum numerical accuracy	Compute + memory latency	CPU
FP16 (half)	Higher throughput, same models	Memory bandwidth	GPU
INT8 (quantised)	Maximal efficiency	SRAM capacity + compiler	NPU

Below you can find the charts with the benchmark results, where

• Various precisions are combined in the charts with bars.
• Devices are mentioned explicitly for the CPU and GPU tests, as the enclosure’s thermal behaviour influences performance

Cinebench 2024
Cinebench 2024 measures sustained CPU and GPU throughput using long-running, floating-point-heavy 3D rendering workloads, exposing the impact of thread scaling, memory bandwidth, and thermal constraints on real-world performance.

Below you can find the Cinebench 2024 CPU and GPU results. Cinebench results are strongly influenced by thermal conditions; however, the benchmark charts above do not disclose the device, enclosure, or cooling configuration used for each M-series processor.

Cinebench 2024 CPU rendering emphasises sustained floating-point throughput, thread scalability, and thermal stability under continuous load:

Cinebench 2024 GPU rendering highlights ray-tracing performance, memory bandwidth, and architectural efficiency under prolonged, thermally constrained workloads:

 

[hifivoice]

Adobe Lightroom Export

This benchmark measures the time required to export 1,000 × 61 MP RAW files from Adobe Lightroom, based on an extensive and consistently executed test by ArtIsRight. Because it represents a real production workload rather than a synthetic test, it provides a clear view of how different Apple silicon configurations perform in a common professional photography task.

Lightroom export is a mixed workload that combines multi-core CPU processing, partial GPU acceleration, sustained memory bandwidth usage, and SSD write throughput. As a result, performance reflects overall system balance rather than any single component in isolation.

The benchmark results from ArtIsRight are shown below.

The results have been reformatted to match the conventions used throughout this benchmark overview. Values are expressed in seconds, where lower values indicate better performance.

Observations and conclusions

1. Limited scaling within the M4 Max tier
The performance difference between a Mac Studio M4 Max with 14 CPU / 32 GPU cores and the 16 CPU / 40 GPU core configuration is relatively small. This suggests that Lightroom export does not scale linearly with additional CPU or GPU cores at this level. Instead, other system components—such as SSD write throughput, memory bandwidth, or pipeline synchronisation—begin to dominate overall performance.

2. Diminishing returns of higher-tier silicon
The Mac Studio M4 Max delivers export times comparable to those of the M3 Ultra. In this workload, the additional cores and resources of the Ultra variant do not translate into meaningful gains, indicating that investing in a higher-tier SoC offers limited benefit for Lightroom export specifically.

3. Strong enclosure and thermal effects
MacBook-based systems consistently show significantly slower export times compared to desktop configurations with similar or even lower-tier M-series processors. This points to thermal constraints playing a major role during sustained export workloads, where performance is limited by cooling capacity rather than peak compute capability.

4. System balance matters more than SoC tier
This effect becomes even clearer when comparing a desktop system with a lower-tier M processor to a laptop with a higher-tier variant. In several cases, the desktop system completes the export faster, despite having fewer cores. This underlines that blindly upgrading to a higher-tier M processor does not guarantee better real-world performance. Overall system design—including cooling, memory configuration, and SSD throughput—must be considered.

5. Clear generational improvements
Across all form factors, a consistent improvement can be observed with each new M-series generation. Architectural refinements continue to reduce export times, confirming that generational progress remains a reliable predictor of performance gains for this workload.

Taken together, these results show that Lightroom export performance is governed by sustained system balance rather than peak specifications, making it a strong counterexample to purely synthetic benchmark rankings.

 

[hifivoice]

Adobe Lightroom Preview

This benchmark measures the time required to generate 1,000 previews from 61 MP RAW files in Adobe Lightroom. The data is again sourced from an extensive and consistently executed test by ArtIsRight, using identical input data and workflow across a wide range of Apple silicon systems.

Unlike export, preview generation is a front-loaded, interactive workload. It is representative of the first phase of many real Lightroom sessions: importing images, browsing, culling, and making initial selections. As such, it provides insight into how responsive a system feels during everyday photographic workflows, rather than how fast it completes a long batch job.

The benchmark results from ArtIsRight are shown below.

As with the export benchmark, the results have been reformatted to match the conventions used throughout this benchmark overview:

Values are expressed in seconds, and lower values indicate better performance.


Observations and conclusions


1. Huge impact of including GPU-based processing
The new Lightroom version 14.5.1 now also uses GPU based processing to create previews. The effect of that is quite stunning, with gains of more than a factor of 2.

As the Macbook M5 Pro was introduced when this new version of Lightroom 14.5.1. has been shipped, and it is only tested with the new version of Lightroom. An estimated version (with less opacity in the chart) has been made to mimic the performance with an old version of Lightroom, and to make the M5 comparable to other samples in the chart.

2. Clear generational improvements
Across all form factors, a consistent improvement can be observed with each new M-series generation. Architectural refinements continue to reduce export times, confirming that generational progress remains a reliable predictor of performance gains for this workload.

3. Less clear differentiation between Pro and Max variations
Although there is a clear benefit amongst the variations, the differentiation between Pro and Max variations is less obvious and intermingled.

There is a strong correlation with the amount of CPU cores in these devices, that with the exception of the M3 series strongly overlap in the amount of CPU cores:

 

[hifivoice]

Finally, a table showing the major specifications of the Apple M1-M5 base processors:

 

[Ursadorable]

Thank you for taking the time to compile these, and sharing them!