Apple's ambitious self-driving car project, code-named Titan, never reached the roads. After nearly a decade of development and billions of dollars in investment, the program was officially canceled in early 2024. However, the project's legacy lives on in a way that few could have predicted. The intense computational demands of autonomous driving pushed Apple's silicon engineers to develop a dedicated neural processing unit, which later became the Neural Engine. This component is now the cornerstone of Apple's AI hardware strategy, powering everything from iPhones to Macs and, soon, dedicated servers.
The Birth of the Neural Engine
The original car project required massive on-device AI processing to handle real-time sensor fusion, object recognition, and path planning. Offloading data to the cloud was impractical due to latency and privacy concerns. Apple's chip architects realized they needed a specialized processor capable of running machine learning models efficiently. This led to the development of the Neural Engine, a dedicated block of circuitry optimized for neural network computations. The first iteration debuted in September 2017 with the A11 Bionic chip inside the iPhone X. At the time, it could perform 600 billion operations per second and was primarily used for computer vision tasks like Face ID, Animoji, and augmented reality filters.
Over successive generations, the Neural Engine grew more powerful. The A12 Bionic doubled the core count from two to eight, increasing performance to 5 trillion operations per second. The A13 Bionic boosted that to 6 trillion, and the A14 Bionic reached 11 trillion. Each iteration brought improvements in energy efficiency and new capabilities, such as real-time language translation and advanced camera computational photography. Yet the Neural Engine remained relatively underutilized in software, as Apple's AI applications lagged behind competitors like Google and Microsoft. The hardware, however, was already setting the stage for a dramatic shift in computing.
Transition to Mac and the M-Series
Apple's transition from Intel processors to its own Apple Silicon, beginning with the M1 chip in 2020, brought the Neural Engine to the Mac for the first time. The M1's 16-core Neural Engine could handle 11 trillion operations per second, enabling features like real-time speech recognition, background blur in video calls, and intelligent photo editing. More importantly, it demonstrated Apple's commitment to on-device AI processing as a core differentiator. The unified memory architecture allowed the Neural Engine to access the same pool of RAM as the CPU and GPU, reducing data movement and improving latency—a critical advantage for AI workloads.
The M1 Pro, M1 Max, and M1 Ultra expanded on this foundation, with the Ultra effectively combining two M1 Max dies to double performance. Neural Engine capabilities remained consistent across the lineup, but the sheer computational power available in the Ultra version made it suitable for even larger machine learning models. Apple's strategy of scaling the Neural Engine across all devices—from the Apple Watch to the Mac Pro—ensured that developers could write once and run everywhere, a key selling point for the ecosystem.
Accelerating Development: M5, M6, and the Leap to M7
In 2023 and 2024, Apple released the M3 and M4 chips, each with incremental Neural Engine improvements. The M3 series featured a 16-core Neural Engine capable of 18 trillion operations per second, while the M4—first introduced in the iPad Pro—bumped that to 38 trillion. These numbers were impressive, but Apple's AI software efforts, such as the delayed launch of Apple Intelligence and limited generative AI features, continued to lag behind industry leaders. Meanwhile, competitors like Qualcomm and AMD were catching up with their own neural processing units, and Nvidia's GPUs dominated cloud AI workloads.
According to reports from a well-known Apple analyst, the company decided to skip the Pro, Max, and Ultra variants of its upcoming M6 chip. This unusual move signaled a deeper shift in Apple's chip roadmap. Instead of incremental upgrades across multiple tiers, Apple is focusing its resources on the M7 generation, which is expected to arrive in the first half of 2027. The M7 will feature significant upgrades to the Neural Engine, likely doubling or tripling its performance compared to the M4. But the most surprising revelation is the M7 Ultra variant, which is being designed to support up to 1.5 terabytes of unified RAM.
This massive memory capacity is unprecedented in consumer chips and indicates that Apple is building the M7 Ultra for server applications. By pairing the Neural Engine with over a terabyte of fast, unified memory, Apple can run large language models and other generative AI workloads entirely on-device, reducing dependence on cloud infrastructure. This aligns with Apple's privacy-focused narrative: sensitive data never leaves the user's device, and even complex AI tasks can be performed locally at high speed.
Implications for Apple's AI Future
The M7 Ultra's server-class capabilities suggest that Apple plans to enter the AI cloud market in a meaningful way, either by offering its own cloud services powered by Apple Silicon or by enabling developers to deploy AI workloads on Apple hardware. The unified memory architecture gives Apple an edge over traditional GPUs, which require separate VRAM and often suffer from bandwidth bottlenecks. Apple's approach allows the Neural Engine, CPU, and GPU to share the same pool of high-bandwidth memory, drastically reducing the time needed to move data between processing units.
For consumers, the M7 and its successors mean that future Macs and possibly iPads will be able to run the most advanced AI models locally—think real-time video generation, conversational agents with voice, and complex data analysis—all without an internet connection. This could transform workflows in creative industries, healthcare, scientific research, and education. The privacy benefits are equally compelling: personal data remains on the device, and users retain control over their information.
Apple's AI hardware leadership, born from the ashes of the car project, now positions the company as a serious contender in the AI arms race. While Google and Microsoft have invested heavily in cloud-based AI, Apple is betting on a hybrid approach: powerful on-device processing supplemented by optional cloud services. The M7 Ultra's massive RAM and enhanced Neural Engine will be the backbone of this strategy, allowing Apple to offer competitive AI features without compromising its core privacy principles.
Other companies are also racing to develop similar on-device AI processors. Qualcomm's Snapdragon X Elite features a Neural Processing Unit with up to 45 trillion operations per second, and Intel's upcoming Lunar Lake chips include an NPU for AI acceleration. However, Apple's advantage lies in its vertical integration: it controls the hardware, the operating system, and the application ecosystem. This allows Apple to optimize the entire stack for AI performance, much like it did with the original iPhone. The Neural Engine's evolution from a car project afterthought to the heart of Apple's AI future is a testament to the company's long-term vision and its willingness to invest in breakthrough technologies, even when the immediate product fails.
The M7 generation is expected to set new benchmarks for on-device AI, and if Apple can couple it with compelling software features, it may finally close the gap with rivals in generative AI. The company's commitment to privacy, combined with increasingly powerful hardware, could redefine what users expect from their devices. Meanwhile, the legacy of Apple's car program continues to drive innovation in unexpected ways, proving that even failed projects can sow the seeds of transformative technologies.
Source: The Verge News