Essay No. 079  ·  Apple Silicon / TSMC / Local AI
Apple Silicon Apple M2 Apple M5 M-Series TSMC Unified Memory AI PC GPU Architecture Semiconductors Local AI

Apple M2 Was the Costly Transition Chip. M5 Shows Apple's Real AI Silicon Direction. M2 M3 M4 M5 A15 IP LPDDR5 Unified memory GPU Neural Accelerators TSMC 3nm Fusion Architecture Local AI

M2 widened the platform with more die area, LPDDR5, GPU and media capability. M5 shows where Apple Silicon was really going: neural acceleration inside the GPU, unified memory bandwidth, local AI, pro workflows, and advanced packaging.

PM
PUGALENTHI MAGENDRAN
May 27, 2026  ·  Research memo  ·  Updating a 2022 Apple M2 die-shot thesis
16 MIN
Thesis
The 2022 article was right that M2 was not Apple's major architectural leap. It was a costly transition chip: larger die, A15-derived CPU IP, more GPU area, larger caches, LPDDR5 support, and higher memory-controller cost. But the 2026 update is that M2 was only the awkward middle step. M3 reset Apple's GPU architecture on 3nm. M4 pushed stronger on-device AI. M5 moved neural acceleration into every GPU core and scaled unified memory bandwidth for local LLMs, diffusion models, graphics, and pro workflows. Apple's real silicon direction is not just faster CPU cores. It is GPU-driven AI, unified memory bandwidth, media acceleration, and advanced packaging riding TSMC's process cadence.
Executive summary
  • In 2022, Apple marketed M2 as a broad upgrade: 20 billion transistors, 24GB LPDDR5 support, 100 GB/s memory bandwidth, 18% faster CPU, 35% faster GPU, and stronger media and Neural Engine capabilities.
  • The die-shot analysis showed a more complicated story: M2 was largely A15-derived IP with a larger die, more GPU area, larger cache and system logic, LPDDR5 support, and a bigger memory controller.
  • The article's key insight was cost: M2 moved from M1's 118.91 mm² die to an estimated scaled 155.25 mm² die, while LPDDR5 and memory-controller area increased platform cost.
  • M3, M4 and M5 clarified Apple's actual direction: GPU architecture, media acceleration, on-device AI, unified memory bandwidth, and advanced packaging mattered more than one huge CPU IPC leap.
  • M5 is the clearest signal: Apple moved neural acceleration into every GPU core and scaled unified memory bandwidth for local AI and pro workflows.

Section 1  ·  Historical frameWhat the 2022 M2 die-shot article got right

The 2022 SemiAnalysis and Locuza piece on M2 made a careful, almost contrarian argument.[1] M2 was not M1.5, but it also was not a fresh architectural leap. M2 was generally based on A15-era IP, just as M1 was broadly based on A14-era IP. M2 was larger and costlier than M1. The authors questioned Apple's published die-image scaling and Locuza rescaled M2 using comparable SRAM and PHY structures to estimate M2 at approximately 155.25 mm² versus M1 at approximately 118.91 mm², with Apple's presented die size implying approximately 141.7 mm². The CPU IP looked close to A15 Avalanche and Blizzard. CPU gains came mostly from clocks and modest IPC movement, while GPU, media, and memory improvements were more meaningful. The article also flagged no clear SRAM shrink from first-generation to second-generation N5 and a notable LPDDR5 memory controller and PHY area increase.[1]

2022 PDF page references used in this essay
  • Apple's M2 marketing panel: 20B transistors, 24GB LPDDR5, 100 GB/s memory bandwidth, 18% faster CPU, 35% faster GPU, 40% faster Neural Engine, and high-performance media engine.
  • M1, M2 and A15 die comparison with Locuza annotations.
  • Apple chip specs table (A13, A14, A15, M1, M2) with die size and transistor density, including the scaled M2 estimate of ~155.25 mm².
  • M1 Firestorm, A15 Avalanche, and M2 Avalanche-derived P-core comparison; M2 P-core estimated at ~2.76 mm².
  • M1 / A15 / M2 E-core comparison and GPU core comparison.
  • SLC comparison and memory controller / PHY comparison, including LPDDR5 area increase to ~14 mm² vs ~8.1 mm² for M1's LPDDR4X.
  • Cost pressure from larger die plus more expensive LPDDR5-6400 vs LPDDR4X-4266.

M2 was not Apple flexing a huge CPU leap. It was Apple paying more silicon and memory cost to widen the platform.

Section 2  ·  Platform wideningNot a CPU revolution

M2's CPU story was modest compared with the GPU, memory, and media story.[1] M2's P-core was based on A15's Avalanche lineage with Mac-specific modifications. M2's E-core looked very close to A15's Blizzard E-core. The P-core uplift came mostly from clocks, not a large IPC jump. The GPU moved to more cores and more shared logic. The media engine and LPDDR5 support mattered more for real Mac workflows than raw CPU IPC. Apple did not spend the M2 die budget on a new CPU generation; it spent it on platform width.

Apple did not spend the M2 die budget on a new CPU generation. It spent it on platform width.

Section 3  ·  Die sizeThe cost story

The page 5 specs table puts the die-size story in numbers. M1 die size is reported at approximately 118.91 mm². The Locuza-rescaled M2 estimate is approximately 155.25 mm². Apple-presented M2 scaling implies approximately 141.7 mm². The rescaling uses identical SRAM cells and PHY structures, and the measurement has an error window because it relies on Apple-provided imagery.[1]

M1 die size
~ 118.91 mm²
Reported M1 die size in the 2022 article's table.
M2 scaled estimate
~ 155.25 mm²
Locuza rescaling using identical SRAM cells and PHY structures.
Apple-presented M2
~ 141.7 mm²
Implied by Apple's published die imagery in the article.
Implication
Fewer dies / wafer
Larger die on the same broad process family means higher per-die cost.

M2's performance story was modest. Its cost story was not.

Section 4  ·  CoresThe P-core and E-core evidence

The page 6 P-core comparison places M2's Avalanche-derived P-core alongside M1's Firestorm and A15's Avalanche. The M2 P-core is estimated at approximately 2.76 mm² in the visual and is larger than M1 and A15 P-core comparisons. The shared L2 cache grew from 12 MB to 16 MB. The page 8 E-core comparison argues the M2 E-core looks nearly identical to A15's Blizzard E-core, and page 9 says the E-core complex grew only modestly compared with the broader CPU complex.[1] The evidence points to continuity, not reinvention.

The CPU evidence points to continuity, not reinvention.

Section 5  ·  GPU and mediaThe more meaningful story

The page 9 GPU comparison shows M1, A15, and M2 GPU cores at similar size per core, but Apple increased the core count, raised GPU clocks, and added more GPU area. The larger media engine improved creative workflows. The article explicitly framed Apple M-series silicon as excellent for creative professionals and Adobe-style workflows.[1] M2 looked underwhelming if judged only by CPU uplift; it made more sense as a creator and bandwidth platform update.

M2 looked underwhelming if judged only by CPU uplift. It made more sense as a creator and bandwidth platform update.

Section 6  ·  MemoryLPDDR5 and unified memory cost

The page 11 SLC and memory-controller comparison reports that each 2 MB SLC data array looked generally the same size across M1, A15, and scaled M2, with no clear SRAM shrink from first-generation to second-generation N5. The memory controller and PHY area grew significantly to support LPDDR5-6400. The article estimates the 128-bit LPDDR5 bus at about 14 mm² versus about 8.1 mm² for M1's 128-bit LPDDR4X, and notes that LPDDR5-6400 was significantly more expensive than LPDDR4X-4266 at the time.[1] Apple paid silicon area and BOM cost for bandwidth. That spend looks different by 2026, when local AI and pro workflows put a real premium on memory bandwidth.

The memory system looked expensive in 2022. By 2026, memory bandwidth looked like the right place to spend.

Section 7  ·  M3The real GPU architecture reset

The M3 family announcement framed M3 around the biggest leap forward in graphics architecture for Apple Silicon, built on 3nm and introducing Dynamic Caching, hardware-accelerated ray tracing, and hardware-accelerated mesh shading, with Dynamic Caching allocating local memory in hardware in real time to improve GPU utilization.[2] M3 was the point where Apple's Mac silicon story shifted from CPU efficiency miracle to GPU, memory, and media architecture.

M3 was the point where Apple's Mac silicon story shifted from CPU efficiency miracle to GPU, memory and media architecture.

Section 8  ·  M4On-device AI became unavoidable

Apple introduced M4 on second-generation 3nm technology, with a Neural Engine capable of up to 38 trillion operations per second and a GPU built on M3's architecture with Dynamic Caching, hardware ray tracing, and mesh shading.[3] M4 Pro and M4 Max carried second-generation 3nm into Macs, with Apple framing the family around Apple Intelligence, on-device AI, and pro workflows.[4] Apple's AI strategy after M4 became platform-wide rather than NPU-only.

M4 made Apple's AI strategy obvious: accelerate the whole device, not just one NPU benchmark.

Section 9  ·  M5The real AI silicon direction

Apple's M5 announcement frames M5 as the next big leap in AI performance for Apple Silicon. M5 is built on third-generation 3nm, has a 10-core GPU with a Neural Accelerator in each GPU core, a 16-core Neural Engine, and unified memory bandwidth of approximately 153 GB/s. Apple claims over 4x peak GPU compute performance for AI versus M4 and up to 45% higher graphics performance versus M4, with explicit framing around local AI workloads including diffusion models and LLMs.[5]

Process
3nm gen 3
Built on TSMC third-generation 3nm technology, per Apple.
GPU AI compute
> 4x vs M4
Peak GPU compute for AI versus M4, per Apple's M5 claims.
Graphics
up to 45%
Graphics performance gain vs M4, per Apple.
Unified memory bandwidth
~ 153 GB/s
Unified memory bandwidth in M5, per Apple's M5 page.

M5 is Apple's answer to the AI PC: not a discrete accelerator card, but neural acceleration inside the GPU plus unified memory bandwidth.

Section 10  ·  M5 Pro and M5 MaxPackaging and bandwidth strategy

Apple's M5 Pro and M5 Max announcement describes a Fusion Architecture that connects two third-generation 3nm dies into a single SoC using advanced packaging, with M5 Pro supporting up to approximately 307 GB/s unified memory bandwidth and M5 Max supporting up to approximately 614 GB/s with up to 128 GB of unified memory. Apple links the memory bandwidth to complex scenes, massive datasets, and higher LLM token generation.[6] The technical specs page is the source of truth for exact configurations.[7]

Fusion Architecture
2 dies
Two third-generation 3nm dies fused via advanced packaging, per Apple.
M5 Pro bandwidth
up to 307 GB/s
Unified memory bandwidth on M5 Pro, per Apple.
M5 Max bandwidth
up to 614 GB/s
Unified memory bandwidth on M5 Max, per Apple.
M5 Max memory
up to 128 GB
Unified memory capacity on M5 Max, per Apple's technical specs.

Apple's pro-chip strategy is becoming packaging plus bandwidth plus local AI.

Section 11  ·  TSMCThe hidden center

Apple's M-series roadmap is tightly coupled to TSMC's process cadence. TSMC's 3nm family includes N3, N3E, and N3P, positioned around improved power, performance, and density, with Fab 18 as the main 3nm production facility.[8] TSMC's smartphone-platform material also frames N3P as part of the broader 3nm cadence supporting mobile and HPC products.[9] Apple's M3, M4, and M5 progression should be framed as Apple architecture riding TSMC node cadence, with Apple as an early premium customer.

Apple Silicon is Apple design riding TSMC's node cadence better than almost anyone else.

Section 12  ·  ASMLIndirectly relevant

ASML matters because EUV enables the advanced nodes Apple uses through TSMC. But the direct evidence for Apple Silicon is Apple's product announcements and TSMC's process roadmap. This essay does not pivot into ASML lithography. The cleaner foundry lens is TSMC process cadence and Apple's ability to be an early premium customer on each new node.

ASML makes the machines. TSMC turns the machines into nodes. Apple turns those nodes into products.

Section 13  ·  The arcThe real Apple Silicon arc

M2 -> M3 -> M4 -> M5 -> M5 Pro / Max  ·  the real Apple Silicon arc
M2 (2022)
5nm-family, A15-derived IP, LPDDR5; wider platform, larger die, more GPU / media / bandwidth. Costly transition chip.
M3 (2023)
3nm with Dynamic Caching, hardware ray tracing, and mesh shading. GPU architecture reset.
M4 (2024)
Second-generation 3nm; stronger Neural Engine (up to 38 TOPS), Apple Intelligence positioning. Platform-wide on-device AI.
M5 (2025)
Third-generation 3nm; Neural Accelerator in each GPU core, 16-core Neural Engine, ~153 GB/s unified memory. Local AI through GPU plus unified memory.
M5 Pro / Max (2026)
Third-generation 3nm + Fusion Architecture; up to ~307 / ~614 GB/s, up to 128 GB unified memory. Packaging and bandwidth for pro AI workflows.

Section 14  ·  Proof pointsWhat Apple must prove now

Apple's open proof points after M5
  1. Can GPU Neural Accelerators become broadly useful through Metal and Core ML?
  2. Can developers easily target Neural Engine, GPU, and CPU without fragmentation?
  3. Can unified memory bandwidth remain a practical advantage for local LLMs and diffusion models?
  4. Can Apple scale memory capacity without destroying product margins?
  5. Can Fusion Architecture scale beyond Pro / Max without thermal and cost problems?
  6. Can Apple keep riding TSMC node cadence as advanced nodes become more expensive?
  7. Can Apple Intelligence make local AI acceleration valuable enough for ordinary users?
  8. Can Apple maintain battery life while adding more AI and GPU compute?
  9. Can Apple's local AI story compete with cloud AI where model size matters more than device efficiency?

Apple's next silicon challenge is not only faster chips. It is making local AI useful enough that the hardware matters.

Section 15  ·  Then and now2022 thesis vs 2026 reality

2022 M2 observation 2026 interpretation Lesson
M2 used A15-derived IP M2 was a transition chip, not a fresh architectural leap Apple was pacing Mac silicon with mobile IP reuse
M2 die looked larger and costlier Later chips used more advanced nodes and packaging to justify cost Silicon area must buy platform capability
CPU uplift looked modest Apple's bigger gains moved to GPU, media, memory, and AI CPU is no longer the whole story
LPDDR5 controller was expensive Unified memory bandwidth became central to local AI and pro workflows Bandwidth was the right long-term bet
No clear SRAM shrink on N5 family Memory hierarchy became harder and more important Scaling is uneven even for Apple
M2 felt awkward M3, M4, and M5 clarified the roadmap M2 was the bridge, not the destination

Section 16  ·  EvidenceEvidence ledger

Claim
Evidence
Interpretation
M2 was marketed as a broad next-generation step
Page 2 of the PDF shows Apple's M2 claims: 20B transistors, 24GB LPDDR5, 100 GB/s memory bandwidth, 18% faster CPU, 35% faster GPU.
Apple framed M2 as a platform upgrade.
The die-size and cost story was the real concern
Page 5 estimates scaled M2 at ~155.25 mm² vs M1 at ~118.91 mm², with Apple-presented M2 at ~141.7 mm².
M2 likely carried a meaningful die-cost increase.
M2 reused A15-era CPU IP
Pages 6 to 9 compare M2 P-cores and E-cores with A15 Avalanche and Blizzard blocks.
M2 was not a fresh CPU architecture leap.
The memory system became more expensive
Pages 11 to 12 discuss no SRAM shrink and a larger LPDDR5 controller / PHY area.
M2 paid area and BOM cost for bandwidth.
M3 was the graphics reset
Apple says M3 introduced Dynamic Caching, hardware ray tracing, and mesh shading on 3nm.
Apple's bigger post-M2 leap was GPU architecture.
M4 pushed stronger on-device AI
Apple says M4 used second-generation 3nm and a Neural Engine up to 38 TOPS.
Apple moved deeper into platform AI acceleration.
M5 moved neural acceleration into the GPU
Apple says M5 has a Neural Accelerator in each GPU core and over 4x GPU AI compute vs M4.
Apple's AI strategy expanded beyond the Neural Engine.
M5 Pro / Max scale with packaging and bandwidth
Apple says Fusion Architecture connects two third-generation 3nm dies; M5 Max reaches up to ~614 GB/s unified memory bandwidth and up to 128 GB unified memory.
Apple is scaling AI / pro workflows through packaging and unified memory.
TSMC's 3nm family underpins the roadmap
TSMC says N3, N3E, and N3P extend 3nm for power, performance, and density.
Apple's roadmap depends heavily on TSMC process cadence.

Section 17  ·  Risk registerRisks and limitations

This essay is an analysis of public disclosures and historical context. It is not investment advice. The honest risks against the read above run in several directions.

The 2022 M2 die-shot analysis used Apple-provided imagery and Locuza rescaling. Estimates are directional, not lab-quality measurements.
Apple performance claims for M3, M4, M5, and M5 Pro / Max are Apple claims. Independent benchmarks may differ.
M5 is not a data-center AI chip. Comparisons with discrete GPUs and accelerators should account for workload scope and memory capacity.
Local AI is one part of the AI story. Many workloads still depend on cloud AI services where model size matters more than device efficiency.
Fusion Architecture economics depend on advanced packaging cost, yield, and tooling. Pro / Max scaling is not free.
Unified memory bandwidth competes with GPU memory capacity and external accelerator strategies. The right balance depends on workload mix.
TSMC node cadence economics get harder at sub-3nm. Apple's premium customer position can compress as nodes get more expensive.
Developers may not target every Apple acceleration path equally. NPU, GPU, and CPU fragmentation can blunt theoretical gains.
Reports about Apple silicon team dynamics should be treated as reported, not as confirmed organizational facts.
Apple Intelligence adoption is still evolving. The hardware case for local AI ultimately depends on whether ordinary users feel the difference.

Section 18  ·  Bottom lineBottom line

Bottom line

The 2022 article was right that M2 was not Apple's major architectural leap. It was a costly transition chip: larger die, A15-derived CPU IP, more GPU area, larger caches, LPDDR5 support, and higher memory-controller cost.

But the 2026 update is that M2 was only the awkward middle step. M3 reset Apple's GPU architecture on 3nm. M4 pushed stronger on-device AI. M5 moved neural acceleration into every GPU core and scaled unified memory bandwidth for local LLMs, diffusion models, graphics, and pro workflows.

Apple's real silicon direction is not just faster CPU cores. It is GPU-driven AI, unified memory bandwidth, media acceleration, and advanced packaging riding TSMC's process cadence.

M2 widened the platform. M5 shows where Apple Silicon was going.

Section 19  ·  DefinitionsGlossary

Apple Silicon
Apple's family of in-house Arm-based SoCs, including A-series for iPhone / iPad and M-series for Mac.
M2
Apple's 2022 Mac SoC. A 5nm-family transition chip with A15-derived CPU IP, larger die, LPDDR5, and a wider GPU / media platform.
M3
Apple's 2023 Mac SoC on 3nm. Introduced Dynamic Caching, hardware ray tracing, and mesh shading.
M4
Apple's 2024 Mac / iPad SoC on second-generation 3nm with a Neural Engine up to 38 TOPS and Apple Intelligence framing.
M5
Apple's 2025 Mac SoC on third-generation 3nm. Adds a Neural Accelerator in each GPU core and ~153 GB/s unified memory.
A15
Apple's 2021 iPhone SoC with Avalanche P-cores and Blizzard E-cores. M2's CPU IP is closely related.
Avalanche
Apple's A15 P-core microarchitecture, reused in modified form in M2.
Blizzard
Apple's A15 E-core microarchitecture, closely reflected in M2's E-cores.
P-core
Performance core. Larger Apple CPU core designed for peak per-thread performance.
E-core
Efficiency core. Smaller Apple CPU core designed for power efficiency in background and parallel tasks.
SLC
System-level cache. A shared cache structure across CPU, GPU, and other blocks in Apple's SoCs.
LPDDR5
Low-Power Double Data Rate 5 memory. Higher bandwidth and capacity than LPDDR4X. M2 adopted LPDDR5-6400.
Unified memory
Apple's architecture in which CPU, GPU, Neural Engine, and other blocks share one memory pool, avoiding explicit copies between devices.
Memory bandwidth
The rate at which data moves between memory and on-chip blocks. A primary constraint on local LLM and pro media performance.
Neural Engine
Apple's dedicated NPU for neural-network inference. M4 introduced a 38 TOPS Neural Engine; M5 has a 16-core Neural Engine.
Neural Accelerator
An AI-acceleration block placed inside each GPU core in M5. Brings neural compute closer to graphics workloads.
GPU compute
General-purpose computation on the GPU, including neural inference workloads. M5 emphasizes GPU compute for AI.
Dynamic Caching
Apple's GPU technology introduced with M3 that allocates local memory in hardware in real time to improve GPU utilization.
Ray tracing
A graphics technique that simulates light paths. M3 introduced hardware-accelerated ray tracing on Apple Silicon.
Mesh shading
A modern graphics pipeline feature for more flexible geometry processing. M3 introduced hardware-accelerated mesh shading on Apple Silicon.
Fusion Architecture
Apple's approach in M5 Pro and M5 Max for connecting two third-generation 3nm dies into a single SoC using advanced packaging.
Advanced packaging
A family of techniques (2.5D, 3D, fan-out, hybrid bonding) used to integrate multiple dies in one package with high-bandwidth, low-power interconnect.
TSMC N3
TSMC's first-generation 3nm process node, used in early 3nm products including A17 Pro and Apple M3.
TSMC N3E
A more cost / yield-tuned variant of TSMC's 3nm node, used in mainstream 3nm products.
TSMC N3P
An evolution of N3E offering improved power, performance, and density characteristics, used by Apple's third-generation 3nm products.
Local AI
AI inference running on a device rather than in the cloud. Benefits from unified memory bandwidth and on-chip accelerators.
Diffusion model
A class of generative AI models used for image generation. Memory-bandwidth and matrix-throughput intensive.
LLM token generation
The rate at which a large language model produces output tokens, often bottlenecked by memory bandwidth on local devices.

Section 20  ·  MethodSources and method notes

How this essay reads sources

The 2022 SemiAnalysis and Locuza M2 die-shot analysis is treated as historical context for the M2 platform widening argument, the page 5 die-size table (with the noted measurement caveats), the pages 6-9 P-core / E-core / GPU comparisons, and the pages 11-12 SLC, memory controller, and cost discussion. Apple's M3, M4, M5, and M5 Pro / Max claims are treated as Apple claims rather than independently validated figures.

The 2026 read is built primarily from Apple's M3 family announcement, the M4 announcement, the M4 Pro / M4 Max announcement, the M5 announcement, the M5 Pro / M5 Max announcement, and the M5 Pro / M5 Max technical specs page, plus TSMC's 3nm and advanced-smartphone-technology pages. ASML is referenced only as indirect context. The structural arguments that M2 was a transition chip rather than a failure, that M3 was the GPU architecture reset, that M4 made on-device AI platform-wide, and that M5 plus Fusion Architecture defines Apple's local-AI direction are independent analysis.

Footnotes  ·  primary sources

  1. SemiAnalysis and Locuza, “Apple M2 Die Shot and Architecture Analysis — Big Cost Increase And A15 Based IP,” 2022 (PDF supplied by author). Historical anchor used in this essay for the page 2 Apple M2 marketing claims (20B transistors, 24GB LPDDR5, 100 GB/s memory bandwidth, 18% faster CPU, 35% faster GPU, 40% faster Neural Engine), the page 4 M1 / M2 / A15 die comparison, the page 5 Apple chip specs table with M1 at ~118.91 mm² and scaled M2 at ~155.25 mm² vs Apple-presented ~141.7 mm², the page 6 Firestorm / Avalanche / M2 P-core comparison and ~2.76 mm² P-core estimate, the pages 8-9 E-core and GPU core comparisons, the page 11 SLC and memory controller / PHY discussion (LPDDR5 bus area ~14 mm² vs ~8.1 mm² LPDDR4X), and the page 12 cost-pressure discussion (LPDDR5-6400 vs LPDDR4X-4266).
  2. Apple, “Apple Unveils M3, M3 Pro, and M3 Max, the Most Advanced Chips for a Personal Computer,” apple.com/…/m3-m3-pro-m3-max. Source for M3 on 3nm, Dynamic Caching, hardware-accelerated ray tracing, hardware-accelerated mesh shading, the “biggest leap forward in Apple Silicon graphics” framing, faster CPU / efficiency cores, and improved media engine.
  3. Apple, “Apple Introduces M4 Chip,” apple.com/…/m4-chip. Source for M4 on second-generation 3nm, the Neural Engine up to 38 trillion operations per second, the M3-derived GPU architecture with Dynamic Caching, hardware ray tracing and mesh shading, and the on-device AI positioning.
  4. Apple, “Apple Introduces M4 Pro and M4 Max,” apple.com/…/m4-pro-m4-max. Source for the Mac family transition to second-generation 3nm, the M4 Pro and M4 Max platform scaling, the Apple Intelligence and pro workflow positioning, and the unified memory framing.
  5. Apple, “Apple Unleashes M5, the Next Big Leap in AI Performance for Apple Silicon,” apple.com/…/m5. Source for M5 on third-generation 3nm, the Neural Accelerator in each GPU core, the over 4x peak GPU compute for AI vs M4, the up to 45% higher graphics performance vs M4, the 153 GB/s unified memory bandwidth, the 16-core Neural Engine, and the local LLM and diffusion model framing.
  6. Apple, “Apple Debuts M5 Pro and M5 Max to Supercharge the Most Demanding Pro Workflows,” apple.com/…/m5-pro-m5-max. Source for Fusion Architecture connecting two third-generation 3nm dies into one SoC, advanced packaging, M5 Pro up to ~307 GB/s unified memory bandwidth, M5 Max up to ~614 GB/s unified memory bandwidth and up to 128 GB unified memory, and the LLM token generation, massive dataset, and complex scene framing.
  7. Apple, “M5 Pro / M5 Max Technical Specifications,” support.apple.com/…/m5-pro-m5-max-tech-specs. Source for exact M5 Pro / M5 Max configurations, memory bandwidth options, and unified memory capacity details used in this essay.
  8. TSMC, “3nm Technology,” tsmc.com/…/3nm. Source for the TSMC 3nm family including N3, N3E, and N3P, the 3nm performance / power / density positioning, and the Fab 18 as main 3nm production facility context.
  9. TSMC, “Advanced Smartphone Technology Platform,” tsmc.com/…/advanced-smartphone-technology. Source for the broader TSMC platform context covering smartphone, HPC, and mobile process cadence, used in this essay only as supporting framing for Apple's TSMC dependence.
← All essays