M5 Pro & M5 Max In-Depth Review: The First Shot of the Chiplet Era

> "Apple has finally stopped making one big monolithic chip. The M5 Pro and M5 Max share the same CPU Tile; only the GPU Tile differs. This isn't a compromise — it's precision engineering."

In March 2026, Apple released the Pro and Max versions of the M5 series, marketed under a "Fusion Architecture." For the first time in Apple Silicon history, the high-end chips are no longer a single piece of silicon:

• M5 Pro and M5 Max share an identical CPU Tile — 18-core CPU, 16-core NPU, Thunderbolt 5 controller, media engine, and SSD controller, all on one tile.
• The only difference is the GPU Tile: 20 GPU cores (M5 Pro) vs. 40 GPU cores (M5 Max).
• They are bonded via SoIC-MH (System on Integrated Chips – Micro-bump Hybrid bonding) packaging.

This is a deliberate engineering decision about yield, thermal management, performance scheduling, and cost — not cost-cutting.

Key points

Why chiplets?

• At advanced nodes (3nm and below), larger die area means exponentially worse yield due to random wafer defects. The reticle limit also caps monolithic die size.
• Unlike M3 Ultra (two complete SoCs stitched together, which is redundant), M5 Pro/Max split functional tiles: one shared CPU Tile + differentiated GPU Tiles.
• Benefits per the analysis (citing Creative Strategies / Ben Bajarin): halved R&D and validation cost for the CPU Tile; GPU bins can be salvaged from defective 40-core tiles (fused down to 32/20 cores); physical separation decouples CPU and GPU thermals, so sustained combined workloads no longer throttle each other.
• SoIC-MH packaging costs more, but the savings in front-end waste more than compensate: "Apple spends more money in the back-end of manufacturing in exchange for massive savings in the front-end."

CPU: two-tier core design, no more Efficiency Cores

| Chip | Super cores | Performance cores | Efficiency cores | GPU | |------|-------------|-------------------|------------------|-----| | M5 | 4 | 0 | 6 | 10 | | M5 Pro | 6 | 12 | 0 | 20 | | M5 Max | 6 | 12 | 0 | 40 | | M4 Max | 12 (then called P-cores) | 0 | 4 | 40 |

• Verified via sysctl/powermetrics (Ars Technica, Andrew Cunningham): Performance cores run up to 4,308 MHz vs. 4,608 MHz for Super cores — only 300 MHz lower — with 128KB L1I / 64KB L1D and 8MB L2 per cluster (vs. 6MB for old E-cores).
• Apple describes them as a new design "derived from the Super Core architecture, optimized for power-sensitive multi-threaded workloads" — analogous to AMD's Zen 4c/5c approach rather than Intel's heterogeneous cores.
• Despite having no E-cores, idle power is below 2W in light desktop use; system idle dropped from 7.6W (M4 Max) to 7.1W, and battery life gained one hour over M4 Max.

CPU benchmarks

| Test | M5 Max | M4 Max | M3 Ultra | Delta | |------|--------|--------|----------|-------| | Geekbench 6 single-core | 4,268 | 3,895 | 3,082 | +10% | | Geekbench 6 multi-core | 29,233 | 25,984 | 27,157 | +12% | | Cinebench 2024 single | 738 | 676 | 573 | +9% | | Cinebench 2024 multi | 8,413 | 7,829 | 12,082 | +7% |

• M5 Max's multi-core score beats the M3 Ultra desktop workstation chip in a laptop.
• Power behavior: ~66W peak (Geekbench), ~50W sustained (Cinebench); in Handbrake the Super core cluster holds 4.24 GHz for 10s, then settles around 3.9 GHz, while the Performance core clusters sustain 4.2–4.3 GHz — the P-cores are the true sustained-workhorse.

GPU: M5 Max ≈ RTX 5070 mobile at ~1/3 the power

| Test | M5 Max (40c) vs M4 Max | vs RTX 5070 mobile | M5 Pro (20c) | |------|------------------------|--------------------|--------------| | 3DMark Steel Nomad | +8% | ahead | 41% slower than M5 Max | | Geekbench Metal / OpenCL | +20–26% | OpenCL: slightly ahead of RTX 5070 | — | | Cinebench 2024 GPU | +40% | — | +46% vs M4 Pro | | Blender | slightly below M4 Max | — | +30% vs M4 Pro |

• M5 Pro's GPU lands between RTX 5050 and RTX 5060 mobile; it wins synthetic benchmarks against Radeon 8060S but loses in real games (driver/API optimization favors DirectX/Vulkan).
• Gaming is improved (+8–24% vs M4 Max in Cyberpunk 2077, Assassin's Creed Shadows, Baldur's Gate 3 at 1080p) but not a strength.

Creative apps (PugetBench)

| Device | Photoshop | Premiere Pro | DaVinci Resolve | |--------|-----------|--------------|-----------------| | MBP 16" M5 Max | 15,875 | 157,049 | 127,090 | | MBP 16" M5 Pro | 15,271 | 105,296 | 83,560 | | Asus ProArt P16 (RTX 5090) | 10,096 | 107,130 | 85,114 | | Asus ROG Z13 (Radeon 8060S) | 11,932 | 57,481 | 53,737 |

M5 Pro nearly matches the RTX 5090 mobile in Premiere Pro; M5 Max crushes every Windows laptop, even in the thermally limited 14" chassis (149,151).

Memory bandwidth

| Spec | M5 Pro | M5 Max | |------|--------|--------| | Bandwidth | 307 GB/s | 614 GB/s | | Max capacity | 48GB | 128GB | | Type | LPDDR5X, 8,533 MHz | LPDDR5X, 8,533 MHz | | Bus width | 256-bit | 512-bit |

614 GB/s suffices to run a quantized 70B-parameter LLM locally (~12 tokens/s in llama.cpp Q4_K_M). M5 Max's memory controllers sit on the GPU Tile, explaining the doubled bandwidth and larger memory ceiling.

AI / Neural Engine

• Every GPU core integrates a Neural Accelerator, giving M5 Max ~4x the AI compute peak of M4 (measured ~19.9 TFLOPS FP16 vs. M4's 15.8).
• Testing with GPT-5.4 (Codex) suggests the ANE behaves like a fast but fixed-shape dense compute engine: excellent on large FP16 matmuls (prefill), much weaker on small, dynamic, token-by-token decode — where memory bandwidth is the bottleneck.

14" vs 16" chassis

The 14-inch MacBook Pro cannot fully unleash M5 Max: GPU power sags from 72W to ~44W sustained, costing 10–18% in Adobe/DaVinci workloads. M5 Pro (38W GPU peak, identical CPU Tile) is arguably the sweet spot for the 14-inch body.

Competition

| Chip | GB6 single | GB6 multi | Peak power | |------|-----------|-----------|------------| | M5 Max | 4,268 | 29,233 | ~30W | | Intel Core Ultra 9 285K | 3,294 | 22,760 | 253W | | AMD Ryzen 9 9950X | 3,251 | 24,026 | 170W |

M5 Max posts the highest consumer single-core score ever, ~30% ahead of desktop flagships at 1/5–1/8 the power. GPU-wise it trades blows with the RTX 5070 (mobile and desktop) at ~72W vs 250W TGP, and beats Radeon 8060S (Strix Halo) by ~15%. Per Notebookcheck, even the best compact Windows machines (Asus ProArt PX13 + Strix Halo) are no match for M5 chips in Photoshop, Premiere Pro, and DaVinci Resolve.

Who should buy

| User | Recommendation | Reason | |------|---------------|--------| | Video editors / colorists | M5 Max + 16" | Dominant PugetBench scores, needs the cooling | | Developers / data science | M5 Pro + 14" | Same CPU as Max, cheaper, portable | | Local AI/ML | M5 Max 128GB | Runs 70B quantized models locally | | Photographers / designers | M5 Pro | Photoshop score near Max | | Gamers | Neither | Windows + RTX 5070 is better | | Students / light office | Base M5 | Pro/Max unnecessary |

The bigger picture

• Apple adopting chiplets at 3nm signals that even the best monolithic designer can no longer avoid chiplets — due to reticle limits, yield economics, and SKU flexibility.
• M5 Ultra (expected WWDC 2026) is rumored at 40,000+ Geekbench 6 multi-core, ~400,000+ Metal score, and 256GB memory, via two M5 Max GPU Tiles.
• The core takeaway: Apple is no longer maximizing percentage gains per generation but system-level efficiency — manufacturing efficiency (chiplets), scheduling efficiency (Super + Performance cores), and compute density (per-core NPUs). Starting with M5 Pro/Max, Apple Silicon's competition has shifted from raw benchmarks to systems engineering.

Sources

• Ars Technica M5 Max review (2026-03-09): https://arstechnica.com/gadgets/2026/03/testing-apples-2026-16-inch-macbook-pro-m5-max-and-its-new-performance-cores/
• Creative Strategies chiplet analysis (2026-03-09): https://creativestrategies.com/research/m5-max-chiplets-thermals-and-performance-per-watt/
• Notebookcheck CPU analysis (2026-03): https://www.notebookcheck.net/Apple-M5-Pro-Max-CPU-Analysis-M5-Max-is-not-much-faster-than-the-M4-Max.1253941.0.html
• Notebookcheck GPU analysis (2026-03-10): https://www.notebookcheck.net/Apple-M5-Pro-M5-Max-GPU-Analysis-M5-Max-GPU-on-par-with-the-GeForce-RTX-5070-and-faster-than-Strix-Halo.1246060.0.html
• Tech Insider benchmark roundup (2026-04): https://tech-insider.org/apple-m5-chip-benchmarks-a-new-standard-for-personal-computing/
• NanoReview GPU comparison (2026-04-16): https://nanoreview.net/en/gpu-compare/geforce-rtx-5070-vs-apple-m5-max-gpu-40-core
• Notebookcheck M5 Max specs (2026-04-25): https://www.notebookcheck.net/Apple-M5-Max-Processor-Benchmarks-and-Specs.1244918.0.html