The launch of the Apple M5 chip is not just a generational update. It is a technological leap that continues to challenge the fundamental dominance of x86 architecture (Intel/AMD). The M5 pushes the boundaries of power-efficient performance, particularly in the emerging field of Artificial Intelligence (AI).
Apple M5 technological deep dive: Cache, Tensors, and the 3nm advantage
The M5 chip, produced on the advanced TSMC 3nm (N3P) process, is fundamentally an “AI-first” chip, focusing on deep integration and specialized acceleration.
Neural Accelerators: The AI powerhouse
The most significant upgrade for the M5 is the integration of advanced Neural Accelerators within the GPU and CPU cores. These accelerators are Apple’s custom, dedicated matrix cores (similar to Nvidia’s Tensor Cores).
- Impact: AI models rely on massive matrix calculations. These specialized cores perform these tasks in parallel on-chip, resulting in preliminary speedups of 3x to 4x in key Large Language Model (LLM) workflows compared to the previous M-series architecture.
The importance of cache on efficiency
Apple has significantly increased the size of the on-chip SRAM cache (L2 and System Level Cache) in the M5.
Impact: By providing the processing cores with extremely fast, on-die memory, the M5 dramatically reduces the number of trips it must make to the slower, power-hungry unified memory. This results in superior power efficiency and a higher thermal ceiling, allowing the M5’s high-performance (P) cores to safely hit speeds up to 4.61GHz sensors, and the 3nm design process
Performance benchmarks: CPU, Gaming, and AI
The Apple M5 processor is not the only one in the market. The competition is difficult, not only from the established manufacturers like Intel or AMD, but also from Qualcomm.
Core technical specifications
What are the differences between the Apple M5, Intel Core Ultra 9 285H, Qualcomm Snapdragon X2 Elite, and AMD Ryzen AI 9 HX 370?
| Feature | Apple M5 (10-Core) | Snapdragon X2 Elite Extreme | Intel Core Ultra 9 285H | AMD Ryzen AI 9 HX 370 |
| Architecture | ARM (RISC) | ARM (RISC) – Oryon Cores | x86 (CISC) – Arrow Lake-H | x86 (CISC) – Zen 5/5c (Strix Point) |
| Process | 3nm (TSMC N3P) | 3nm (TSMC) | 5nm (Intel 20A equivalent) | 4nm (TSMC) |
| Total Cores / Threads | 10 Cores / 10 Threads | 18 Cores / 18 Threads | 16 Cores / 16 Threads | 12 Cores / 24 Threads |
| Max Frequency (P-Core) | ~4.61GHz | 5.0GHz (Single-core Boost) | 5.4 GHz | 5.1 GHz |
| Base TDP / Max TDP | ~25W} (Estimated) | ~25W (Estimated) | 45W / 115W | 28W / 54W |
CPU performance summary (Geekbench 6.5)
Seeing and comparing numbers is nice, but what do these numbers mean? Please keep in mind that we have estimations so far, but they indicate who the winner is.
| Benchmark | Apple M5 (10-Core) | Snapdragon X2 Elite Extreme | Intel Core Ultra 9 285H | AMD Ryzen AI 9 HX 370 |
| Single-Core | ~ 4,263 (1st) | ~ 4,080 | ~ 2,961 | ~ 3,978 |
| Multi-Core | ~ 17,862 | ~ 23,491 (1st) | ~ 16,661$ | ~ 18,500 (Estimated) |
Gaming & Graphics benchmarks (3DMark)
The integrated GPU performance is a tightly contested area, with both ARM chips significantly outperforming the integrated graphics of the mobile x86 competition. These scores are based on the 3DMark Unlimited stress tests.
| Benchmark (FPS) | Apple M5 (10-Core GPU) | Snapdragon X2 Elite Extreme (Adreno) | Intel Core Ultra 9 285H (Arc iGPU) | AMD Ryzen AI 9 HX 370 (Radeon iGPU) |
| Wild Life Extreme | ~ 73 FPS (1st) | ~ 69 FPS | ~ 42 FPS (Estimated) | ~ 45 FPS (Estimated) |
| Solar Bay (Ray Tracing) | ~ 90.4 FPS | ~ 88.05 FPS | ~ 30 FPS (Estimated) | ~ 40 FPS (Estimated) |
| Steel Nomad Light | ~ 39.1 FPS | ~ 41.69 FPS (1st) | N/A | N/A |
GPU Verdict: The Apple M5 takes the overall lead in most GPU benchmarks, particularly the intensive Wild Life Extreme and the Ray Tracing-focused Solar Bay, showcasing the power of its MetalFX upscaling technology. However, the Snapdragon X2 Elite Extreme achieves a narrow victory in the new Steel Nomad Light test, indicating a competitive and well-optimized Adreno GPU, making it a strong alternative for future Windows gaming.
AI performance and application acceleration
AI performance is no longer just about the Neural Processing Unit (NPU). Modern workloads are distributed across the CPU, GPU, and NPU.
| Metric | Apple M5 (16-Core N.E.) | Snapdragon X2 Elite Extreme | Intel/AMD |
| NPU Power (Peak TOPS) | ~ 57 TOPS | ~ 80 TOPS (1st) | ~ 13-50 TOPS (NPU Only) |
| Geekbench AI 1.5 Score | ~ 57,242 | ~ 88,615 (1st) | N/A (Low NPU scores) |
| Adobe Photoshop (AI Filters) | “Instant” (Optimized for Neural Engine/Metal) | “Very Fast” (Optimized for Hexagon/Windows Studio) | “Fast” (Relies more on CPU/GPU) |
The AI Battle: The Snapdragon X2 Elite Extreme offers a significantly higher theoretical peak NPU power (80 TOPS), which directly translates into a decisive win in the standardized Geekbench AI benchmark. This is a clear lead in raw, sustained AI computation.
Real-World Software: Apple’s advantage lies in its deep software integration (Apple Intelligence tools and frameworks like Core ML and Metal Performance Shaders). In applications like Adobe Photoshop, which have been heavily optimized to use Apple’s Neural Engine, the performance for features like Generative Fill or Neural Filters feels instantaneous, even if the raw TOPS number is lower than Qualcomm’s. The X2 Elite is also rapidly improving its ecosystem through its Hexagon NPU and partnership with Microsoft’s AI platform.
Why do we have this difference in performance? ARM vs. x86 architecture explained
The stunning single-core performance and efficiency gap are not accidental. They are the result of Apple successfully leveraging a fundamentally different, more efficient approach to processor design.
Instruction Set: RISC (ARM) vs. CISC (x86)
The core difference lies in how the chips interpret code:
| Feature | Apple ARM (RISC) | x86 (CISC) |
| Philosophy | Reduced Instruction Set Computing | Complex Instruction Set Computing |
| The Advantage | Simple, fixed-length instructions require less power and fewer transistors to decode, making the core inherently smaller and more power-efficient, leading to higher efficiency and better single-core IPC. | Complex, variable-length instructions consume significantly more energy and generate more heat due to the complex internal logic needed for decoding. |
System Integration: System-on-a-Chip (SoC) vs. Discrete Components
The M5 is an integrated powerhouse, while x86 systems rely on external pathways:
- Apple Silicon (SoC): The M5 is a System-on-a-Chip. All major components (CPU, GPU, Neural Engine, and memory controller) are integrated onto a single package. This physical proximity allows data to move faster and with far less power loss.
- AMD/Intel (Dedicated): Traditional PCs use dedicated components. The CPU and (often) the GPU are separate chips connected by slower, power-hungry motherboard pathways like the PCIe bus, creating bandwidth bottlenecks.
3. Unified Memory Architecture (UMA)
Apple’s use of Unified Memory is the greatest advantage for data-intensive professional workflows:
- Apple’s UMA: The CPU and GPU share a single, high-bandwidth pool of RAM that is physically integrated onto the chip package. Data transfer between the CPU and GPU is virtually instantaneous because there is no need for time-consuming and energy-intensive data copying between separate memory pools.
- x86’s Separate Memory: PCs must copy data from the CPU’s system RAM to the dedicated GPU’s VRAM over the system bus, which is a significant bottleneck for applications dealing with large data sets (e.g., video production, massive local AI models).
Is the M5 good enough?
The M5 solidifies Apple’s lead in the critical market for high-performance, low-power computing. While rivals like the Snapdragon X2 Elite Extreme and AMD Zen 6, push higher core counts and dedicated NPU power to win the multi-core and AI TOPS crowns, Apple’s architectural choices, RISC, SoC design, and Unified Memory, provide a significant, sustained advantage in performance per watt and graphics efficiency, making the M5 the chip of choice for the modern, mobile, and AI-centric era.
