Apple Silicon & Local NVMe Surprise: MacBook Neo Outperforms Cloud Servers in Key Database Test

Apple’s latest entry-level MacBook Neo, powered by the M4 chip and equipped with a 512GB NVMe SSD, has delivered an unexpected performance victory in a high-stakes database workload comparison conducted by DuckDB’s data engineer Gábor Szárnyas. In tests against two cloud instances—one with 16 AMD EPYC vCPU cores and 32GB RAM, the other with 192 Graviton4 vCPUs and 384GB RAM—the MacBook Neo demonstrated superior efficiency in cold-run scenarios and competitive median performance in hot runs, challenging long-held assumptions about the necessity of cloud-based data processing.

Why the MacBook Neo’s Local Storage Beats Cloud Attached Disks in Cold Tests

The ClickBench Benchmark Setup

Szárnyas put the 512GB MacBook Neo through its paces using two industry-standard benchmarks: ClickBench and TPC-DS. ClickBench evaluates database performance through 43 queries focused on aggregation and filtering operations, executed on a single wide table containing 100 million rows. The dataset consumes approximately 14GB when stored in Parquet format and swells to 75GB in CSV—a critical detail, as file size directly impacts read times and memory usage during query execution.

The benchmark was run in two phases: a cold run, where caches are cleared to simulate first-time access, and a hot run, where the system leverages cached data for faster execution. The results revealed a dramatic disparity. In the cold run, the MacBook Neo completed all ClickBench queries in under one minute—up to 2.8 times faster than both the c6a.4xlarge (16 vCPU, 32GB RAM) and the c8g.metal-48xl (192 vCPU, 384GB RAM). This outcome underscores the latency penalty associated with cloud-based storage systems, which rely on network-attached disks that introduce significant overhead during data retrieval.

Szárnyas highlighted the underlying reason for this performance gap: 'The cloud instances have network-attached disks, and accessing the database on these dominates the overall query runtimes,' he wrote in his blog post. 'The MacBook Neo has a local NVMe SSD, which is far from best-in-class, but still provides relatively quick access on the first read.' Even a modest NVMe drive offers orders-of-magnitude lower latency than most cloud storage solutions, particularly when handling large datasets that exceed available RAM.

Hot Run Results Subvert Expectations

The narrative shifted during the hot run, where the c8g.metal-48xl surged ahead with a blistering 4.35-second completion time, followed by the c6a.4xlarge at 47.86 seconds. The MacBook Neo, while still competitive, lagged behind at 54.27 seconds—a time that, while slower than the cloud servers, represents only a 10% increase from its cold-run performance. Crucially, the laptop’s median query runtime remained superior to the mid-tier cloud instance (c6a.4xlarge), and its total run time was only 13% slower despite the cloud box boasting 10 additional CPU threads and quadruple the RAM.

This outcome suggests that while cloud servers excel in sustained, cached workloads thanks to their vast computational resources, the MacBook Neo’s Apple Silicon M4 chip and efficient local storage architecture deliver exceptional performance in scenarios where data locality and low latency are paramount—such as initial data ingestion, ad-hoc queries, or environments with limited network bandwidth.

TPC-DS: Surviving the Scaling Challenge with Grace

The TPC-DS benchmark, a more complex and widely respected standard for evaluating database performance, presented a greater challenge. TPC-DS includes 24 tables and 99 queries, many of which incorporate advanced SQL features like window functions, correlated subqueries, and multi-table joins. Szárnyas ran the test at two scale factors (SF): SF100 and SF300, representing dataset sizes of 100GB and 300GB, respectively.

Strong Performance at SF100

At SF100, the MacBook Neo handled the workload with remarkable efficiency. It breezed through most queries, achieving a median runtime of 1.63 seconds and completing the entire benchmark in just 15.5 minutes. This performance placed it on par with or ahead of many cloud-based systems operating under similar constraints, highlighting the raw efficiency of Apple’s unified memory architecture and the M4 chip’s energy-efficient yet powerful core design.

Memory Constraints Emerge at SF300

The SF300 test, however, exposed the MacBook Neo’s 512GB storage limitation. While the median query runtime remained solid at 6.90 seconds, the system frequently spilled data to disk—using up to 80GB of space for temporary overflow. This memory constraint manifested in prolonged execution times for several queries, most notably Query 67, which took a staggering 51 minutes to complete. Despite these challenges, the MacBook Neo ultimately persevered, finishing the full benchmark in 79 minutes.

Szárnyas noted the resilience of the hardware and software ecosystem working in tandem: 'While the laptop’s total runtime is only about 13% slower than the cloud box despite having far fewer CPU threads and RAM, it’s clear that memory capacity remains the primary bottleneck.' This observation underscores a critical trade-off in local computing: while cloud servers can scale memory and CPU resources dynamically, local systems are constrained by fixed hardware configurations but benefit from unparalleled data locality and reduced latency.

A History of Apple Silicon Pushing Boundaries in Database Workloads

This isn’t the first time Apple’s in-house silicon has demonstrated surprising capabilities in unconventional workloads. Earlier this year, Szárnyas and his team ran the TPC-H benchmark on an iPhone 16 Pro—equipped with the A19 Pro chip—under extreme conditions: submerged in a bucket of dry ice at -50°C. Despite the thermal stress, the device completed the benchmark in 478.2 seconds, proving the robustness and thermal efficiency of Apple’s chip design. These results collectively challenge the assumption that high-performance computing (HPC) and database workloads are exclusively the domain of x86 servers or cloud-based infrastructure.

Key Takeaways: What These Benchmarks Really Mean

• The MacBook Neo outperforms mid-tier cloud servers in cold-run database scenarios due to superior local NVMe SSD latency, completing ClickBench queries up to 2.8x faster than cloud instances.
• Despite limited RAM (just a fraction of cloud server configurations), the MacBook Neo achieves competitive median query performance and total runtimes only 13% slower than a 192-core cloud server in hot runs.
• TPC-DS results at SF100 show the MacBook Neo completing complex 99-query workloads in 15.5 minutes, with median runtimes of 1.63 seconds—highlighting Apple Silicon’s efficiency in real-world applications.
• Memory constraints become evident at SF300, where the 512GB MacBook Neo spills data to disk and faces extended query times, though it ultimately completes the benchmark in 79 minutes—demonstrating resilience.
• These findings suggest a viable future for local database processing on consumer-grade laptops, particularly for developers, small businesses, and edge computing use cases where cloud costs, latency, and network dependence are prohibitive.

The Broader Implications: Rethinking Local vs. Cloud Data Processing

The DuckDB benchmark results carry significant implications for the future of data infrastructure. For years, the prevailing wisdom has dictated that large-scale database operations—especially those involving complex analytical queries—belong in the cloud, where elastic resources and massive parallelism can be summoned on demand. However, the MacBook Neo’s performance calls this paradigm into question, particularly in contexts where real-time processing, data privacy, or cost efficiency are priorities.

Small businesses, indie developers, and researchers often operate under tight budgets and bandwidth constraints. For these users, the ability to run meaningful database workloads locally on an affordable laptop could democratize data analysis, reduce cloud dependency, and improve responsiveness. Moreover, the rise of edge computing—where data is processed closer to its source to minimize latency—aligns naturally with the strengths of local hardware like the MacBook Neo.

Yet, it’s important to temper expectations. While the MacBook Neo punches above its weight in certain scenarios, its limitations are undeniable. The SF300 TPC-DS test revealed that when datasets grow beyond available RAM, even the M4 chip struggles with disk spills and extended runtimes. This highlights the need for a balanced approach: leveraging local hardware for development, prototyping, and lightweight analytics, while reserving cloud resources for heavy-duty, memory-intensive workloads.

The Role of DuckDB in Modern Database Benchmarking

DuckDB, the lightweight analytical database engine used in these benchmarks, has emerged as a critical tool for evaluating modern hardware performance. Unlike traditional database systems designed for server environments, DuckDB is optimized for embedded use cases—running directly within applications and leveraging the host system’s resources without the overhead of client-server architectures. This makes it an ideal benchmarking tool for consumer-grade hardware like the MacBook Neo.

Szárnyas, who serves as DuckDB’s Engineering Lead, has become a prominent voice in the data engineering community for his rigorous, reproducible benchmarking methodologies. His work not only highlights hardware capabilities but also influences database engine development, pushing frameworks like DuckDB to optimize for newer architectures such as Apple Silicon.

What This Means for Consumers and Developers

For consumers considering the MacBook Neo as a primary machine, these results suggest that the laptop is far more capable than its $1,099 price tag might imply—at least for certain workloads. Developers prototyping database applications, running Jupyter notebooks with embedded analytics, or performing ad-hoc data exploration can do so with confidence, knowing that the M4 chip and local storage deliver real-world performance that rivals cloud servers in many common scenarios.

However, professionals handling large-scale analytical workloads, real-time transaction processing, or multi-user database systems will still benefit from cloud or on-premises server solutions. The MacBook Neo shines in isolation—ideal for a solo developer or a small team with intermittent, moderate-scale data needs—but is not a replacement for enterprise-grade infrastructure.

The Future: Can Local Hardware Replace Cloud in Data Workloads?

While the MacBook Neo’s performance is impressive, it’s unlikely to signal the end of cloud computing for database workloads. Instead, it points toward a more hybrid future—one where local hardware handles initial development, testing, and lightweight production, while cloud resources scale up for heavy lifting. The rise of ARM-based chips like Apple Silicon, combined with advancements in local storage and memory technologies, is eroding the traditional advantages of cloud-only processing.

Moreover, environmental concerns are pushing companies to reduce their carbon footprint. Local computing, when powered by efficient chips like the M4, can significantly lower energy consumption compared to large-scale cloud data centers. As sustainability becomes a higher priority, benchmarks like these may gain even greater relevance, prompting a reevaluation of where and how data is processed.