Apple launches M5 chip: 10 core CPU, 10 core GPU (Neural Accelerators in each) More AI power

Apple announced M5 chip today, marking the company’s most significant AI-focused chip redesign since the introduction of Apple Silicon. Built on third-generation 3nm technology, M5 introduces a fundamental architectural change: every GPU core now includes a dedicated Neural Accelerator, delivering over 4x the peak GPU compute performance for AI workloads compared to M4. We saw this first happening in new Apple iPhone 17 chip

The chip powers the updated 14-inch MacBook Pro , iPad Pro, and Apple Vision Pro, all available for pre-order today.

M5 represents Apple’s response to the escalating compute demands of on-device AI, particularly for running large language models locally a capability that’s becoming central to Apple Intelligence and developer workflows.

The Neural Accelerator Architecture: Apple’s AI Bet

M5’s defining feature is its next-generation 10-core GPU architecture with a Neural Accelerator embedded in each core. This isn’t a minor optimization it’s a fundamental rearchitecture of how Apple processes AI workloads.

The performance claims:

• 4x peak GPU compute for AI versus M4
• 6x peak GPU compute for AI versus M1
• 45% graphics performance increase over M4 (with ray tracing)
• 30% faster general graphics performance over M4
• 2.5x faster graphics than M1

The technical angle that matters: Apple is following a similar path to Nvidia’s tensor core strategy, but with tighter integration. Rather than separate AI accelerators, Apple embedded Neural Accelerators directly into GPU cores. This allows the GPU to handle both graphics and AI workloads simultaneously without context switching between different silicon blocks.

For developers, this means applications using Apple’s frameworks Core ML, Metal Performance Shaders, and Metal 4 automatically gain performance improvements. Apps can also directly program the Neural Accelerators using Tensor APIs in Metal 4 for custom AI workloads.

Johny Srouji, Apple’s senior vice president of Hardware Technologies, positioned M5 as the “next big leap in AI performance for Apple silicon,” emphasizing that “every compute block of the chip is optimized for AI.”

CPU, Neural Engine, and Memory: Comprehensive Upgrades

Beyond the GPU overhaul, M5 delivers improvements across the entire system-on-chip:

CPU: Up to 10 cores (four performance cores, six efficiency cores) with what Apple claims is “the world’s fastest performance core.” Multithreaded performance increases up to 15% over M4.

Neural Engine: Enhanced 16-core design provides faster AI inference for on-device Apple Intelligence features and third-party applications using Apple’s Foundation Models framework.

Unified Memory Bandwidth: 153GB/s, a nearly 30% increase over M4’s 120GB/s and more than double M1’s bandwidth. This is critical for AI workloads, which are often memory-bandwidth constrained rather than compute-constrained.

Memory Capacity: Up to 32GB on base M5, supporting larger AI models running entirely on-device.

The unified memory architecture where CPU, GPU, and Neural Engine share a single memory pool becomes increasingly important as AI models grow. Higher bandwidth means faster data access for all processing units, reducing bottlenecks when running large language models or processing high-resolution media.

Real-World AI Performance: Time to First Token

Apple provided specific benchmarks for local LLM performance using an 8-billion parameter model with 4-bit quantization. Testing measured “time to first token” the latency between submitting a prompt and receiving the first response using a 16,000-token prompt.

The results suggest 3.5x faster prompt processing on M5 versus M4. For context, this metric tests how quickly the chip can prefill the key-value cache for transformer models a compute-intensive operation that determines how responsive local AI feels.

Community analysis on platforms like Reddit indicates this performance gain translates to practical usability improvements. With M5’s bandwidth and Neural Accelerators, an M5 Max (expected to have 300GB/s bandwidth) could theoretically handle mixture-of-experts models like QwQ-32B at speeds approaching 2,000-3,000 tokens per second for prompt processing—fast enough for real-time coding assistants and long-form writing tools.

However, technical community responses remain cautious. Time to first token doesn’t measure model loading from storage (a separate bottleneck) and represents ideal-case performance. Real-world results will depend on how frameworks like Ollama and LM Studio optimize for M5’s architecture.

Manufacturing Process: TSMC N3P

M5 uses “third-generation 3-nanometer technology,” which industry sources identify as TSMC’s N3E process node. The M5 Pro and MAx models are expected to ise n3P process. This represents an advancement over M4’s N3E process, offering improved power efficiency and higher transistor density.

TSMC’s 3nm FinFlex technology allows customization of transistor performance characteristics, enabling Apple to optimize different chip blocks for specific workloads. The N3E process likely delivers the power efficiency gains necessary for M5’s increased performance without proportional battery life degradation.

Speculation suggests M5 Pro and Max variants are delayed because they’ll utilize TSMC’s System-on-Integrated-Chips (SoIC) 3D stacking technology, which requires additional qualification time. If accurate, the higher-tier chips could feature even more dramatic performance improvements through chiplet-based designs similar to what Intel is pursuing with Panther Lake’s Foveros packaging.

Graphics and Ray Tracing Improvements

While AI performance dominates the marketing, M5’s graphics capabilities represent a significant generational leap. Apple’s third-generation ray-tracing engine combined with second-generation dynamic caching delivers up to 45% better performance in ray-traced workloads compared to M4.

For Apple Vision Pro, M5 enables 10% more pixels rendered on the micro-OLED displays with refresh rates up to 120Hz, reducing motion blur and improving visual clarity. For gaming and 3D applications on MacBook Pro and iPad Pro, the enhanced shader cores and rearchitected GPU provide smoother gameplay and faster rendering times.

The graphics improvements matter because they demonstrate M5 isn’t just an “AI chip” it’s a comprehensive system-on-chip upgrade that benefits all compute-intensive workloads, from creative applications to gaming to spatial computing.

Storage Performance: 2x Faster SSDs

The new 14-inch MacBook Pro with M5 includes up to 2x faster SSD performance compared to the previous generation, with configurations up to 4TB. This directly impacts AI workflows where loading large language models from storage can be a significant bottleneck.

Faster storage combined with higher memory bandwidth means users can load multi-billion parameter models more quickly and swap between different models with less latency important for workflows that utilize multiple specialized models rather than a single general-purpose LLM.

The Competitive Context: Apple vs. The Field

M5 arrives as the AI chip landscape becomes increasingly competitive. Qualcomm’s Snapdragon X Elite processors target Windows laptops with dedicated AI accelerators. Intel’s Panther Lake, launching in 2026, promises up to 180 Platform TOPS for AI workloads. AMD’s Ryzen AI processors integrate XDNA AI engines for on-device inference.

Apple’s advantage remains vertical integration. By controlling the chip design, operating system, and developer frameworks, Apple can optimize the entire stack for AI performance in ways competitors can’t easily replicate. The Neural Accelerators in M5’s GPU work seamlessly with macOS and iPadOS APIs, providing performance gains automatically for applications using Apple’s frameworks.

However, the lack of M5 Pro and Max variants at launch raises questions. Professional users with M4 Pro or Max machines have no upgrade path yet. The base M5 with 32GB maximum memory and a 10-core GPU serves the mainstream market but won’t satisfy users running the largest AI models or handling multi-stream 8K video workflows.

Apple Intelligence and On-Device AI

M5’s improvements directly benefit Apple Intelligence the company’s suite of on-device AI features. Image Playground, Writing Tools, and other Apple Intelligence capabilities run faster on M5’s Neural Engine and benefit from increased unified memory bandwidth.

More significantly, M5’s architecture supports running larger foundation models entirely on-device. Apple’s Privacy-first approach to AI requires processing sensitive data locally rather than sending it to cloud servers. M5’s combination of Neural Accelerators, Neural Engine, and high-bandwidth unified memory makes this approach more practical for complex AI tasks.

For developers using Apple’s Foundation Models framework, M5 delivers faster inference without requiring code changes. The architectural improvements translate directly to better performance for any application leveraging on-device AI.

The Upgrade Calculus : Should you upgrade

For users considering an upgrade, the math depends on workload:

Upgrade makes sense if:

• Running local LLMs frequently (3.5x faster prompt processing matters)
• Using AI-heavy creative applications (Draw Things, AI video tools)
• Gaming or 3D work benefits from 45% ray tracing improvement
• Current machine is M1 or older (6x AI performance gain is substantial)

Consider waiting if:

• You have M4 and don’t run demanding AI workloads (15% CPU uplift is incremental)
• You need M5 Pro/Max for professional workflows (not available at launch)
• Your workloads are primarily CPU-bound without AI components

Community analysis suggests the performance improvements justify upgrading if the price difference between M4 and M5 models remains around 10% ($100-150 USD), which appears likely based on Apple’s historical pricing.

Apple M5 chip – Gearing towards AI

M5 represents Apple’s most significant AI-focused chip architecture change since introducing Apple Silicon. Neural Accelerators in every GPU core deliver 4x AI performance gains over M4, while improvements to CPU, Neural Engine, and memory bandwidth enhance the entire system. The chip targets the growing importance of on-device AI, particularly local LLM inference, where M5’s architecture provides meaningful real-world performance improvements.

The lack of M5 Pro and Max variants at launch limits appeal for professional users, but for mainstream users running Apple Intelligence, creative AI tools, or experimenting with local language models, M5 delivers a genuine generational leap. Combined with faster storage and higher memory capacity, it makes on-device AI workflows that were marginal on M4 genuinely practical on M5.

Whether this performance translates to sustained competitive advantage depends on how quickly competitors catch up and whether Apple can maintain its architectural lead as AI workloads continue evolving.

For now, M5 sets a new performance bar for on-device AI in consumer devices and positions Apple to extend that lead with the M5 Pro, Max, and Ultra variants expected later this year.