In an interesting test, DuckDB’s Gábor Szárnyas compared the 512GB MacBook Neo to various cloud servers to see how Apple’s new entry-level laptop performs with heavy database workloads. The situation is as follows.
MacBook Neo competes with cloud servers with up to 4x more memory
Blog post titled Big Data on the Cheapest MacBook (via Boing Boing), Szárnyas describes how he benchmarked the MacBook Neo using two benchmarks: ClickBench and TPC-DS.
ClickBench has 43 queries that focus on aggregation and filtering operations. The operation is performed on a single wide table containing 100 million rows and uses approximately 14 GB when serialized to Parquet and 75 GB when saved in CSV format.
TPC-DS has 24 tables and 99 queries, many of which are more complex and include features such as window functions. Although TPC-H has been thoroughly optimized, we believe that the results of TPC-DS are still valuable.
In all tests, the MacBook Neo was pitted against two cloud instances.
- c6a.4xlarge with 16 AMD EPYC vCPU cores and 32 GB RAM.
- c8g.metal-48xl has 192 Graviton4 vCPU cores and 384 GB RAM.
The ClickBench benchmark ran two tests: a cold run, which measures performance when the cache is empty, and a hot run, which measures performance after the system has available cache.
In cold runs, the MacBook Neo significantly outperformed both cloud instances, completing all queries within a minute. This is up to 2.8x faster than comparable products.
While impressive, DuckDB explains:
Of course, there is an explanation for this if you dig deeper into the settings. Cloud instances have network-attached disks, and access to these databases dominates overall query execution time. The MacBook Neo has a local NVMe SSD, which, while far from best in class, provides relatively fast access on first reads.
Things took a turn during the hot run test. c8g.metal-48xl finished running in 4.35 seconds, c6a.4xlarge was runner-up at 47.86 seconds, and MacBook Neo came in last at 54.27 seconds, about 10% faster than the cold run.
However, it’s worth noting that the MacBook Neo can outperform the medium-sized cloud instance, c6a.4xlarge, in terms of median query execution time. And even though the cloud box has 10 more CPU threads and 4x more RAM, the laptop’s total run time is only about 13% slower.
When it comes to TCP-DS benchmarks, DuckDB’s comparative details are less detailed, but we can see that the MacBook Neo still performs very well considering the hardware.
On SF100, the laptop handled most queries with ease, with a median query execution time of 1.63 seconds and a total execution time of 15.5 minutes.
With the SF300, memory constraints started to appear. Median query execution time was still good at 6.90 seconds, but it was clear that DuckDB could use up to 80 GB of space to spill to disk, making some queries take a long time. Most notably, query 67 took 51 minutes to complete. However, the hardware and software worked together tirelessly and eventually passed the test, completing all queries in 79 minutes.
Interestingly, this wasn’t the first time they tested the A19 Pro chip. When the iPhone 16 Pro was released, we placed the device inside a bucket of dry ice at -50°C and ran the TCP-H benchmark in 478.2 seconds.
Click this link to learn more about benchmarking DuckDB on MacBook Neo.
Worth checking out on Amazon
