This is an original hands-on test report—using environments that could realistically occur in production to put this high-memory VPS through the wringer.
Project I: A realistic large-scale compilation task (full Flutter + Rust Android app build) was used to stress-test this machine—no benchmark scripts, just a production-grade workload.
Test machine
| Item | Configuration |
|---|---|
| CPU | Intel Xeon Platinum 8275CL @ 3.00GHz × 12 vCPU |
| Memory | 47 GB (no swap) |
| Disk | 480 GB SSD |
| OS | Debian 12 (bookworm) |
| Location | Hong Kong node |
Compilation performance: this thing really delivers
Using a heavyweight Android project such as NipaPlay/Yozora (Flutter + Rust/Erika core + native NDK layer) as the stress test:
| Metric | Measured result |
|---|---|
| Rust core compilation (12 threads fully loaded) | CPU sustained 100% utilization, no throttling |
| Peak memory usage | 23 GB (gradle 8G + kotlin 2G + cargo freely using 12 concurrent jobs) |
| First full build (3-architecture APK) | 30.5 minutes (including downloading all dependencies) |
| arm64 single-architecture + warm-cache rebuild | 554 seconds (9.2 minutes)  |
| Full pipeline (once the environment is ready) | About 10 minutes to produce a signed APK |
| Multicore SHA256 throughput (openssl) | About 5.5 GB/s |
The same task repeatedly crashed with OOM on local WSL (7.6 GB of memory), but never crashed once after being moved here—the combination of high memory and 12 cores is an overwhelming advantage for resource-hungry gradle/rust/cargo workloads.
Disk performance: excellent sequential speeds, strong 4K performance
Directly measured with fio (--direct=1):
| Item | Measured result |
|---|---|
| Sequential write (1M blocks, iodepth 32) | 1.24 GB/s |
| 4K random read (qd64 × 2 jobs) | 22.2k IOPS (86.9 MiB/s) |
| 4K random write (qd64 × 2 jobs) | 22.3k IOPS (87.0 MiB/s) |
Sequential throughput of 1.2 GB/s means extracting the NDK/Flutter SDK (about 3 GB) takes only a few seconds, and writing build artifacts to disk is absolutely not a bottleneck. 4K performance of 22k IOPS is normal for cloud disks, and small-file reads and writes during compilation caused no noticeable slowdown.
Compilation usage charts—the machine handles huge projects with ease:
Project II: Headless browser concurrency test
The test webpage was
Richify 富有科技 | richify.net (downloaded and run locally)
Richify (high CPU and memory usage warning; opening it is not recommended unless you have a dedicated GPU or at least 16 GB of memory)
I was confident this thing would put the VPS under serious pressure.
The Richify page used in the test looks like a corporate website on the surface, but its source contains Three.js r184, dual-layer Canvas, dynamic particles, code rain, frosted-glass filters, and a huge number of DOM animations. In a headless environment, this type of page places an enormous load on the rendering engine from repaint and compositing operations. Under this workload, running 32 concurrent pages, scrolling each page six times, and taking full-viewport screenshots, the local loopback achieved a P50 of 309 ms and a P95 of just 348 ms, with zero failures. This shows that Chromium’s multi-process parallel rendering pipeline is exceptionally robust.
After setting up a production environment and running 160 concurrent instances directly for a small project, it still failed to put any pressure on the VPS. Chromium PSS peaked at 22.54 GiB.
One-sentence verdict
A compilation monster, blazing sequential disk performance, and strong 4K performance.
12 cores, lots of memory, and a 1.2 GB/s sequential disk—a workstation built for resource-hungry tools such as gradle, cargo, and rustc.
Next, we’ll hand the environment over to Windows for testing.
CPU test:
387 single-core and 4140 multi-core. Let’s see how it compares with a 12th-generation i7 that also has 12 cores.
As expected, it still can’t beat the i7-12650H. After all, the cost of electricity in Hong Kong has to be factored in, along with the overhead of VPS virtualization. Still, I think it should be more than capable of crushing the old E3/E5 “Dufu” servers.
A5P’s disk:
In actual testing, the A5P disk’s performance was indeed inferior to the Seagate gaming drive in my computer.
However, considering that hard-drive prices have risen sharply and QLC drives are not durable at all, VT’s disk is impressively generous at this price.
Finally, let’s look at the performance of the additional 1 TB HDD
Sequential reads reached 1591 MB/s, while sequential writes held steady at 1275 MB/s. That sequential throughput is truly outstanding. This kind of large disk probably wouldn’t be suitable for compilation because of its 4K performance, but it is perfect for media storage, storing and moving video assets, transcoding files, and archiving large projects.
4K read and write speeds were limited to around 20 MB/s. I remember that people once stored small-file databases and similar data on this kind of disk, filling it with junk files and causing cached writes to take two or three full days.
This time, the provider has imposed limits to prevent small junk files from clogging the server’s underlying I/O and to keep the whole machine stable. Hopefully, the A5P will remain stable this time.
Finally, here is the most conventional hardware test:
Summary
With a solid foundation of 12 cores and 47 GB of memory without swap, it can comfortably handle high-pressure concurrent rendering by hundreds of headless browsers, making it ideal for demanding preprocessing tasks in the AI era, including automated Agent data collection, corpus cleaning, and multimodal webpage parsing.
Sequential throughput exceeding 1 GB/s combined with a 1 TB storage-class large-capacity disk means that archiving large AI knowledge-base corpora, keeping a high-concurrency local vector database (Vector DB) running, and maintaining model caches can finally leave the anxiety of OOM crashes behind.
For NodeSeek users seeking the best possible value and serious practical users, this machine, located at a low-latency Hong Kong node, is not only a productivity monster for heavy compilation; with its highly fault-tolerant memory foundation, solid multi-process rendering, and concurrent pipelines, it also serves as an accessible all-purpose foundation for AI data engineering and private Agent and MCP matrices.
