Key Takeaways
- M5 wins outright on single-thread speed and AI work; M4 was the CPU-focused jump; M3 introduced the 3nm era and hardware ray tracing.
- Base memory bandwidth climbed from 100GB/s on M3 to 120GB/s on M4 to 153GB/s on M5 — roughly a 53% gain across two generations.
- Apple credits M5 with more than 4x the peak GPU compute of M4 for AI tasks, because every one of its GPU cores carries a dedicated Neural Accelerator.
- Multi-core gains are modest generation to generation: M5 delivers about 15% higher multithreaded performance than M4 on the base chip.
- In the 14-inch MacBook Pro, the M5 posted a Geekbench 6 single-core score of 4,263 — at the time, the highest recorded for any Mac or PC processor in that database.
- M3 Max topped out at 400GB/s bandwidth and 128GB of memory; M4 Max raised that to 546GB/s; M5 Max reaches 614GB/s.
- The M5 family renamed Apple’s fastest cores “super cores” and, in the Pro and Max tiers, dropped efficiency cores entirely in favour of a new performance core.
- M3 owners gain the most from upgrading. M4 owners gain little on raw CPU throughput but a great deal on local AI and graphics.
M5 microprocessor. Image credit: Apple
Three MacBook chip generations now sit side by side on the used and refurbished market, and they were each built around a different priority. The short answer to which performs best: M5 leads on every axis that matters in 2026, but its advantage over M4 is concentrated in graphics and on-device AI rather than in raw processor throughput. If you compile code, crunch spreadsheets or edit photos, the M4 remains close. If you run diffusion models or language models locally, the gap becomes enormous.
The reason sits in how Apple spent each generation’s transistor budget. M3 moved the line to a 3nm process and gave Mac graphics hardware ray tracing and Dynamic Caching for the first time. M4 poured its gains into the CPU and a much faster Neural Engine. M5 rebuilt the GPU around matrix maths, putting a Neural Accelerator inside every graphics core — which is why its headline numbers are AI numbers.
Silicon at a Glance: M3, M4 and M5 Base Chips
| Specification | M3 (Oct 2023) | M4 (2024) | M5 (Oct 2025) |
|---|---|---|---|
| Process | First-generation 3nm | Second-generation 3nm | Third-generation 3nm |
| CPU configuration | 8-core (4 performance + 4 efficiency) | 10-core (4 performance + 6 efficiency) | 10-core (4 super + 6 efficiency) |
| GPU | Up to 10 cores, first-gen ray tracing | Up to 10 cores, second-gen ray tracing | 10 cores, third-gen ray tracing, Neural Accelerator per core |
| Neural Engine | 16-core, roughly 18 TOPS | 16-core, roughly 38 TOPS | 16-core, revised for higher throughput |
| Memory bandwidth | 100GB/s | 120GB/s | 153GB/s |
| Maximum unified memory | 24GB | 32GB | 32GB |
CPU: Where the Real Gains Landed
M4 was the CPU generation. It added two efficiency cores over M3 and pushed clocks higher, producing roughly a 23% single-core improvement in Geekbench 6 testing and a comfortable lead in multi-core work. That was the year Apple’s chip design team spent its budget on general-purpose compute.
M5 slowed that pace deliberately. Apple states the M5 CPU delivers up to 15% faster multithreaded performance than M4 — respectable, but far from the M3-to-M4 leap. Single-thread performance is a different story. Apple describes the M5’s fastest core as the world’s fastest CPU core, attributing it to wider front-end bandwidth, a rebuilt cache hierarchy and better branch prediction. The benchmark evidence supports the claim: a 14-inch MacBook Pro with M5 recorded 4,263 in Geekbench 6 single-core, the highest figure in that database for any Mac or PC chip at the time, alongside 17,862 multi-core — about 20% ahead of the M4 model it replaced.
A useful reference point for anyone weighing an upgrade: the MacBook Air with M5 scored 17,073 multi-core, roughly 15% above the M4 Air and about 16% faster than a MacBook Pro carrying the older M3 Pro chip. An entry-level 2025 chip now outruns a professional 2023 one.
Graphics and AI: The M5 Architecture Change
The GPU is where the three generations diverge sharply. M3 introduced Dynamic Caching, mesh shading and hardware ray tracing to Apple silicon. M4 refined that architecture. M5 replaced the maths engine underneath it.
Each of the M5’s ten GPU cores contains a dedicated Neural Accelerator — small matrix-multiplication units that handle the linear algebra behind neural networks without the data having to travel to a separate block. Apple puts the result at more than 4x the peak GPU compute of M4 for AI workloads and over 6x that of M1. Graphics performance, helped by a third-generation ray-tracing engine, comes in up to 45% above M4. In the 14-inch MacBook Pro specifically, Apple quotes up to 3.5x the AI performance of the previous M4 model and up to 1.6x the graphics.
Those numbers matter well beyond gaming. Running a language model locally is bounded by two things: how fast the chip can chew through the prompt, and how fast memory can feed weights to the compute units. Neural Accelerators attack the first; the jump to 153GB/s attacks the second. This is the same bottleneck that shapes decisions at data-centre scale, where AI infrastructure sets the ceiling on model performance long before software optimisation does.
Pro and Max Tiers: A Wider Spread
The gap between generations widens considerably once you move above the base chips. The M5 Pro and M5 Max are built on Apple’s Fusion Architecture, which bonds two 3nm dies into one system on a chip using advanced packaging — a structural break from the single-die designs of M3 and M4.
| Chip | CPU cores | GPU cores | Memory bandwidth | Max unified memory |
|---|---|---|---|---|
| M3 Pro | Up to 12 (6P + 6E) | Up to 18 | 150GB/s | 36GB |
| M4 Pro | Up to 14 (10P + 4E) | Up to 20 | 273GB/s | 64GB |
| M5 Pro | 15 or 18 (super + performance) | 16 or 20 | 307GB/s | 64GB |
| M3 Max | 14 or 16 | 30 or 40 | 300 or 400GB/s | 128GB |
| M4 Max | Up to 16 (12P + 4E) | Up to 40 | 546GB/s | 128GB |
| M5 Max | 18 (6 super + 12 performance) | 32 or 40 | 460 or 614GB/s | 128GB |
Note the naming change. Apple retired the performance-and-efficiency split at the Pro and Max tiers, pairing six “super cores” with twelve new performance cores tuned for multithreaded work at lower power. Benchmark results reflect the redesign: an 18-core M5 Max recorded 29,233 multi-core in Geekbench 6, edging past the 32-core M3 Ultra desktop chip and landing roughly 14–15% above the M4 Max. Early Metal graphics scores for the same chip clustered between 218,772 and 232,718.
Which Generation Should You Buy?
Coming from M3, the case for upgrading is straightforward — you gain bandwidth, cores, a redesigned GPU and the entire Neural Accelerator architecture in a single step. Coming from M4, the decision hinges on workload. Video editors, 3D artists and anyone running models on-device will see the difference immediately. Developers whose day is dominated by compilation and test runs will notice far less, and the survey data on how the AI boom has reshaped developer priorities suggests that mix is shifting fast toward the former group.
It is also worth setting expectations against dedicated hardware. A MacBook is remarkably capable for local inference thanks to unified memory, but discrete accelerators still win on batched throughput — a trade-off visible in the pricing of Nvidia’s RTX Spark superchip and in the design philosophy behind purpose-built silicon such as tensor processing units. Apple’s advantage is that the fast memory and the fast maths travel in your bag.
One caveat on the figures above: Apple’s performance multipliers are vendor-measured on selected industry benchmarks and specific configurations, and third-party benchmark scores vary with thermal headroom, memory capacity and software version. Treat them as directional rather than exact.
If you are interested in this topic, we suggest you check our articles:
- How Much Does the Nvidia RTX Spark Superchip Cost?
- What Is a Tensor Processing Unit (TPU)?
- How Big a Role Does AI Infrastructure Play in Model Performance?
- What Is Required for a Premium AI Infrastructure System?
- How the AI Boom Has Shaped Developers’ Interests in the Past 5 Years
Sources: Apple Newsroom (M5), Apple Newsroom (M5 Pro and M5 Max), Apple Tech Specs, Apple MacBook Pro Comparison, MacRumors, Macworld, Notebookcheck
Written by Alius Noreika
