Apple A18 Pro (Mobile NPU 35 TOPS, iPhone 16 Pro)
Overview
Apple A18 Pro is the SoC for Apple's iPhone 16 Pro / Pro Max, released 2024-09-09. Built on TSMC 2nd-gen 3nm (N3E) with a 6-core CPU (2P + 4E), 6-core GPU, 16-core Neural Engine 35 TOPS, and 8GB LPDDR5X unified memory. Performance is 15-20% up over A17 Pro, with Neural Engine +20%. Apple Intelligence on-device LLM (3B parameters) runs entirely on the A18 Pro, making it the industry's most powerful on-device AI chip.
Key positioning:
- A18 (iPhone 16): 6-core CPU (2P+4E), 5-core GPU, 16-core NE 35 TOPS
- A18 Pro (iPhone 16 Pro/Pro Max): 6-core CPU + 6-core GPU + 35 TOPS NE (double Pro GPU)
- A18 Pro vs A17 Pro: CPU +15%, GPU +20%, NE +20%, power efficiency -20%
- Apple Intelligence: 3B LLM on-device inference (iPhone 16 series + M1+ Mac devices only)
Core Specifications
| Item | Parameter |
|---|
| Architecture | Apple Silicon (A18 Pro, mobile SoC) |
| Process | TSMC 2nd-gen 3nm (N3E) |
| Transistors | ~20 billion |
| CPU Cores | 6 (2P + 4E) |
| CPU Max Frequency | 4.05 GHz (P-core, highest mobile CPU frequency in the industry) |
| GPU Cores | 6 (double Pro, vs A18 5-core) |
| Neural Engine | 16-core, 35 TOPS (A17 Pro 35 TOPS, +0% nominal but +20% real multi-task) |
| Unified Memory | 8GB LPDDR5X |
| Memory Bandwidth | 51.2 GB/s |
| Storage | NVMe (128GB / 256GB / 512GB / 1TB) |
| Display Engine | ProMotion 120Hz + Always-On Display |
| Media Engine | Hardware H.264/HEVC/ProRes/AV1 encode + decode |
| ISP | 2nd-gen 48MP + ProRAW |
| TDP | ~8-10W (one of the most power-efficient AI chips) |
| Launch | iPhone 16 Pro / Pro Max 2024-09-20 |
| Price (iPhone 16 Pro 256GB) | $1,099 |
Comparison with A17 Pro
| Metric | A18 Pro (2024-09) | A17 Pro (2023-09) | Improvement |
|---|
| Process | TSMC N3E | TSMC N3 | 2nd-gen 3nm |
| CPU Cores | 6 (2P+4E) | 6 (2P+4E) | Same |
| CPU Frequency | 4.05 GHz | 3.78 GHz | +7% |
| CPU Performance | +15% (single-core) / +20% (multi-core) | Baseline | Significant |
| GPU Cores | 6 | 6 | Same |
| GPU Performance | +20% | Baseline | Significant |
| Ray Tracing Performance | +2x | Baseline | Significant |
| Neural Engine | 35 TOPS | 35 TOPS | Same TOPS, multi-task +20% |
| Memory | 8GB LPDDR5X | 8GB LPDDR5 | Slight upgrade |
| Bandwidth | 51.2 GB/s | 51.2 GB/s | Same |
| Power Efficiency | -20% | Baseline | Significant |
Apple Intelligence On-Device LLM
| Dimension | Spec |
|---|
| Model | Apple Foundation Model 3B (3 billion parameters) |
| Architecture | In-house Transformer (Llama-like) |
| Quantization | 2-bit / 4-bit mixed quantization (2.5 GB model size) |
| Runtime Memory | 2.5 GB (A18 Pro 8GB sufficient) |
| Performance | ~30 tok/s (on-device LLM inference) |
| Latency | < 100ms TTFT |
| Battery Impact | < 5% per hour (basic queries) |
| Offline Operation | Fully offline (no cloud) |
Apple Intelligence key point: 3B model + 2-bit quantization = 2.5GB, A18 Pro 8GB memory fits it, and 30 tok/s on-device inference is the industry's first phone-level LLM real-time generation.
Apple Intelligence Features
| Feature | On-Device/Cloud | Model |
|---|
| Writing Tools (rewrite/proofread/summarize) | On-device 3B | 30 tok/s |
| Image Playground (image generation) | On-device + Cloud Diffusion | SDXL distilled |
| Genmoji (generated emoji) | On-device 3B | < 1s |
| Siri intelligent Q&A | On-device 3B | 30 tok/s |
| Mail / Notification summaries | On-device 3B | Instant |
| Photo recognition (pets / plants) | On-device 3B | Real-time |
| Photo Memories (memory videos) | Cloud Diffusion | Cloud |
| ChatGPT integration | Cloud GPT-4o | Cloud |
| Task | A18 Pro NE | A17 Pro NE | Improvement |
|---|
| Core ML inference | 35 TOPS | 35 TOPS | Same |
| Multi-task parallel | 8x models simultaneous | 6x models | +33% |
| Image recognition (ResNet-50) | < 1ms | 1.2ms | +20% |
| Speech recognition (Whisper 1B) | ~5x real-time | ~4x real-time | +25% |
| Stable Diffusion 1.5 (1 step) | ~0.8 sec | 1.0 sec | +25% |
| LLM 3B inference | 30 tok/s | 25 tok/s | +20% |
A18 Pro 35 TOPS is the strongest NPU in iPhone history and the industry's most powerful mobile AI chip (Qualcomm Snapdragon 8 Gen 3 has 45 TOPS but 7-8W TDP vs A18 Pro 8-10W; Apple NE achieves higher effective utilization).
GPU 6-Core + Hardware Ray Tracing
| Dimension | A18 Pro 6-core | A17 Pro 6-core | Improvement |
|---|
| Core Count | 6 | 6 | Same |
| Performance | +20% | Baseline | Significant |
| Hardware Ray Tracing | +2x | Baseline | Significant |
| Hardware Mesh Shading | +50% | Baseline | Significant |
| AAA Games (e.g. Death Stranding, Resident Evil) | Smooth 30fps | Borderline 30fps | Significant |
| Item | Details |
|---|
| Vendor | Apple Inc. |
| Product | iPhone 16 Pro / iPhone 16 Pro Max |
| Release Date | 2024-09-09 (Apple Wonderlust 2024) |
| Availability | 2024-09-20 |
| Launch Starting Price | iPhone 16 Pro 128GB: $999 / 256GB: $1,099 / 512GB: $1,299 / 1TB: $1,499 |
| Fab | TSMC N3E (TSMC 3nm 2nd-gen) |
| Packaging | TSMC InFO_SoC (Apple co-designed) |
| Target Market | Consumer / premium smartphones |
Apple A18 Pro vs Qualcomm Snapdragon 8 Gen 3
| Metric | Apple A18 Pro | Qualcomm Snapdragon 8 Gen 3 |
|---|
| CPU Cores | 6 (2P+4E) | 8 (1+5+2) |
| CPU Frequency | 4.05 GHz | 3.3 GHz |
| GPU | Apple 6-core | Adreno 750 |
| NPU | 35 TOPS | 45 TOPS |
| TDP | 8-10W | 7-8W |
| Memory | 8GB LPDDR5X | 16GB LPDDR5X |
| Efficiency (CPU perf/W) | Industry-leading | Slightly lower |
| Efficiency (NPU TOPS/W) | 4.4 TOPS/W | 5.6 TOPS/W |
| AI On-Device LLM | Apple Intelligence 3B | Snapdragon AI 2.5B |
| Price | $999+ (iPhone) | $500+ (Android) |
A18 Pro advantages: single-core CPU +30% (4.05 GHz vs 3.3 GHz), Apple Intelligence on-device LLM 3B (SD 8G3 only 2.5B), efficiency optimization (Apple iOS deep integration).
SD 8G3 advantages: NPU 45 TOPS (vs A18 Pro 35 TOPS), memory 16GB (vs A18 Pro 8GB), multi-vendor Android choice.
Use Cases
- iPhone 16 Pro / Pro Max (ProRes video, ProRAW photography, AAA games)
- Apple Intelligence on-device LLM (3B model 30 tok/s)
- iOS 18 intelligence (mail summaries, photo recognition, writing tools)
- AAA gaming (Death Stranding, Assassin's Creed with ray tracing)
- Vision Pro developer (visionOS 2 compatible)
- Not suitable for: datacenters (mobile SoC)
- Not suitable for: Android deployment (Apple devices only)
- Not suitable for: training (inference only)
Apple A-Series Evolution (Recent 4 Generations)
| Generation | Release | Process | CPU Cores | GPU Cores | NE | NPU TOPS | Memory |
|---|
| A15 Bionic | 2021-09 | 5nm | 6 (2+4) | 4-5 | 16-core | 15.8 | 4-6GB |
| A16 Bionic | 2022-09 | 4nm | 6 (2+4) | 5 | 16-core | 17 | 6GB |
| A17 Pro | 2023-09 | 3nm | 6 (2+4) | 6 | 16-core | 35 | 8GB |
| A18 Pro | 2024-09 | 3nm N3E | 6 (2+4) | 6 | 16-core | 35 | 8GB |
A18 Pro key improvements: N3E process (2nd-gen 3nm) + power efficiency -20% + GPU ray tracing +2x + CPU frequency 4.05 GHz (industry's highest).
Key Features
- 35 TOPS NE: strongest NPU in iPhone history
- Apple Intelligence 3B on-device LLM: 30 tok/s
- 2nd-gen 3nm (N3E): power efficiency -20%
- CPU 4.05 GHz: highest mobile CPU frequency
- GPU hardware ray tracing +2x: AAA gaming ray tracing
- 8GB LPDDR5X 51.2 GB/s: sufficient for on-device AI
- Drawbacks: Apple devices only, 8GB memory limit, non-expandable