Provision a server in two different regions, run the same benchmark suite on both, and see how much the numbers drift.
For providers that mix hardware generations across data centers, buy storage capacity from different upstream vendors, or quietly oversell certain regions, the answer is: a lot. You get fast numbers in the region that shows up in their marketing screenshot and slower numbers everywhere else.
That’s not what happened when we tested IP ServerOne for the HostAdvice Web Hosting Awards 2026.
We ran four live servers across Malaysia, Singapore, and Hong Kong. The sequential disk read came back at 238 MiB/s on every single server. Not approximately. Exactly. The CPU scaling remained linear from 2 to 8 cores. The stress tests produced zero failures across all three Linux instances and every stressor we threw at them.
IP ServerOne has been operating managed cloud infrastructure in Malaysia since 2003. The benchmarks suggest they’ve spent those years building the same infrastructure consistently rather than cutting corners region by region. That’s why they earned a Top 25 Cloud Hosting 2026 award.
What We Tested and How
We provisioned four live instances across two product lines and three regions, ran the same benchmark suite used for every provider in the 2026 awards cycle, and submitted a genuine technical support ticket to test the team’s depth.
The four servers were:
- NovaCloud Malaysia, GeneralOpt-C4: 4 vCPU AMD EPYC, 15 GiB RAM, 400 GiB SSD, Ubuntu 24.04, at 0.55 points per hour
- NovaCloud Singapore, RAMOpt-C2: 2 vCPU AMD EPYC, 30 GiB RAM, 400 GiB SSD, Ubuntu 24.04, at 0.61 points per hour
- NovaCloud Hong Kong, GeneralOpt-C4: 4 vCPU AMD EPYC, 15 GiB RAM, 50 GiB SSD, Ubuntu 24.04.4 LTS, Kernel 6.8.0-124-generic
- NovaCloud Hong Kong, GeneralOpt-C8: 8 vCPU AMD EPYC, 30 GiB RAM, 400 GiB SSD, Windows Server 2022, at 0.94 points per hour
- NovaGPU Malaysia, RTX3090-24GB-G1: 1x NVIDIA RTX 3090 with 24 GB VRAM, 8 vCPU AMD EPYC, 117 GiB RAM, Ubuntu 24.04, at 2.15 points per hour
These aren’t identical configurations. The Singapore instance is a RAM-optimized plan with 2 vCPUs instead of the 4-core general-purpose plan in Malaysia.
I chose that deliberately. A real review tests the platform across plan types, not just across regions.
Here’s what the full benchmark suite returned:
| Metric | Result | What It Tells You |
|---|---|---|
| CPU single-thread (Malaysia) | 1,654.64 events/sec | Matched within 19 events across all three Linux servers |
| CPU single-thread (Hong Kong) | 1,610.41 events/sec | Within normal cloud variance of Malaysia and Singapore |
| CPU multi-thread, 4 cores (Malaysia) | 6,624.70 events/sec | 4.00x the single-thread result. Near-perfect linear scaling |
| CPU multi-thread, 4 cores (Hong Kong) | 6,259.75 events/sec | Linear scaling confirmed across a third region |
| CPU multi-thread, 8 cores (NovaGPU) | 13,240.00 events/sec | Exactly 8x. Every core pulling equal weight |
| Memory write throughput | 6,322 – 6,680 MiB/sec | Tight range across Malaysia, Singapore, Hong Kong, and NovaGPU |
| Memory read throughput | 7,394 – 7,527 MiB/sec | Sub-millisecond latency on all four servers |
| Sequential disk read | 238 MiB/sec | Identical on every server in every region |
| Sequential disk write | 212 – 219 MiB/sec | Consistent SSD storage backend confirmed across data centers |
| Random 4K IOPS (read/write) – Malaysia | 8,102 / 8,100 | Symmetric read/write under concurrent load. Strong for databases |
| Random 4K IOPS (Hong Kong) | 2,600 / 2,613 | Lower than other regions, likely storage contention rather than tier difference |
| Network download (Malaysia) | 262 – 265 Mbps | Sub-2ms idle latency; consistent across runs |
| Network download (Hong Kong) | 261 – 262 Mbps | Closest performance to Malaysia across regions |
| Network download (Singapore) | 155 – 157 Mbps | Lower throughput; relevant for SG-targeted workloads |
| Network download/upload (NovaGPU) | 1,010 / 1,045 Mbps | 1 Gbps+ both directions; zero packet loss |
| China connectivity from Hong Kong | 32.3ms avg latency | Stable cross-border latency; minor ICMP loss due to rate limiting |
| Stress test failures | Zero | CPU, memory, and disk stress tests passed across all servers |
| 90-day uptime | 99.97% | Under 43 minutes downtime; no unplanned outages |
| Support first response | 8 minutes | Direct response from a named system engineer |
| Support technical follow-up | 74 minutes | Clarifying question first, then full technical answer |
These results are not just about raw speed. They show that IP ServerOne’s cloud infrastructure behaves consistently across regions, plan types, and sustained-load scenarios.
The Performance Results
The benchmarks across IP ServerOne’s three Linux servers showed that the infrastructure is built the same way across all servers.
1. CPU: It Scales Exactly as It Should
Single-thread CPU performance across Malaysia, Singapore, and the NovaGPU instance landed between 1,646 and 1,666 events per second. That’s a spread of 19 events across three separate servers in two different regions.
Hong Kong added a fourth data point that holds the same story. Single-thread performance came in at 1,610 events per second, slightly below Malaysia and Singapore but within the kind of variance you expect across cloud instances rather than a sign of different hardware. Multi-thread scaling across 4 cores returned 6,259 events per second, maintaining the near-linear pattern seen on every other server in this test.
That kind of consistency tells you IP ServerOne is provisioning AMD EPYC cores from the same hardware generation across data centers, not mixing old and new gear between locations.
Multi-thread scaling held up just as cleanly. The Malaysia C4 returned 6,624 events per second across 4 cores, which is almost exactly 4x the single-thread baseline. The NovaGPU hit 13,240 across 8 cores, essentially 8x.
That’s linear scaling, and linear scaling is the ideal outcome. Every vCPU is working, not one or two carrying the rest.
What does this mean for your workload? If you’re running containerized applications, parallel database queries, CI/CD pipelines, or anything that splits across cores, you’ll get the performance the plan advertises. There’s no padding in those numbers.
2. Storage: 238 MiB/s Everywhere, No Exceptions
This was the result that stood out most across the entire test run.
Sequential disk reads came back at 238 MiB/s on the Malaysia NovaCloud, 238 MiB/s on the Singapore NovaCloud, and 238 MiB/s on the NovaGPU instance. Three servers. Two regions. One number.
The Hong Kong Linux instance returned 239 MiB/s sequential read and 219 MiB/s sequential write, continuing the pattern cleanly. The random 4K result is the one number in this entire test that stands out: 2,600 IOPS on both read and write, roughly a third of what Malaysia and Singapore delivered. Sequential performance on the same server was completely normal, which rules out a bad disk.
That’s a provider running a consistent storage backend rather than mixing disk tiers between locations. If you’ve ever been burned by a hosting provider that advertises NVMe across the board but quietly uses slower storage in certain regions, this is the counterexample.
Random 4K I/O is the number that matters most for production use, because that’s what databases actually do. The Malaysia and Singapore NovaCloud instances both returned over 8,000 IOPS on simultaneous read and write. The read and write numbers were within 2 IOPS of each other on the Malaysia instance.
That kind of symmetry under concurrent load is a strong signal of well-configured underlying storage.
The NovaGPU came in lower at around 6,300 IOPS on random 4K, likely from higher storage contention on the GPU host nodes. If storage throughput is as critical as GPU compute for your workload, that’s worth factoring in.
The stress test sealed the verdict. Disk I/O bogo ops landed at 3,818 to 3,845 per second across all three servers. The three results are within 27 of each other after three minutes of sustained maximum load. Zero failures.
3. Network: Malaysia, Singapore, and a Surprise
Network performance is where the results get more nuanced, and it’s worth walking through region by region because the picture changes significantly.
The Malaysia NovaCloud instances delivered 262 to 265 Mbps download and 272 Mbps upload across two separate test runs against different servers. Both runs landed within 3 Mbps of each other. That’s a real number, not a lucky test window. Idle latency to a Kuala Lumpur server came in under 2ms, and packet loss was zero.
Singapore came in lower at 155 to 157 Mbps. Consistently lower, across both runs. That’s not a deal-breaker in absolute terms, but it’s a gap worth knowing about. If your users are concentrated in Singapore or Southeast Asia, the Malaysia region delivers a noticeably stronger network result at the same plan type. Choose accordingly.
The Hong Kong NovaCloud Linux instance returned 261 to 262 Mbps download and 272 Mbps upload across two runs, making it the closest regional result to Malaysia in the entire test. Both runs landed within 0.5 Mbps of each other. Idle latency sat at 1.96 to 2.00ms, slightly higher than Malaysia but entirely consistent for intra-region Hong Kong testing.
Then there’s the NovaGPU result, which was the standout number in the entire test run.
Over 1 Gbps on both download and upload. Across two separate runs against different servers. With zero packet loss each time. Idle latency under 1ms on the first run.
For AI and machine learning workloads where you’re moving large datasets, pushing model weights across the network, or running distributed training jobs, that bandwidth changes what’s practical. Iteration speed isn’t just about GPU compute. It’s also about how fast you can get data in and results out. IP ServerOne’s NovaGPU handles both.
China Connectivity from Hong Kong
Most APAC providers list Hong Kong as a region. Fewer can tell you what it actually delivers for cross-border traffic to mainland China.
We pinged Alibaba’s mainland DNS server from the Hong Kong instance across 20 packets. Average latency: 32.3ms. Variance across all arriving packets: 0.089ms.
That consistency matters more than the headline number. A network struggling with the cross-border route shows up in jittery latency, not a fixed number that barely moves. What we saw was a stable connection, not a lucky average.
The 20% packet loss is ICMP rate limiting on Alibaba’s end, standard practice for Chinese infrastructure operators. The latency on arriving packets never flinched. For applications that need reliable low-latency access into mainland China, the Hong Kong region delivers it. IP ServerOne also offers a China Premium Route add-on to further optimize that traffic, which we did not test here.
4. Support: We Tried to Make It Hard
We didn’t ask IP ServerOne’s support team whether they offer root access. We asked whether NovaCloud supports custom kernel versions and what the process is for loading a specific kernel module on a Linux instance.
That’s a question with two distinct interpretations. Loading a kernel module typically requires root access. Replacing the kernel entirely depends on the hypervisor configuration. A support team that doesn’t know the difference would either give you a generic non-answer or get the answer wrong.
The ticket was acknowledged in 8 minutes. Not by a generic support inbox, but by Leon Lee Chin Xian, System Engineer. A name and a title.
The follow-up arrived 74 minutes later. It was a clarifying question: which interpretation did we mean? That’s the right move.
A team that asks before answering understands the question well enough to know it has multiple answers. That’s more useful to you than a confident response that addresses the wrong version of your problem.
The knowledge base reinforced that impression. Well-organized, practically arranged, and kept current alongside the product lineup.
One detail that stood out is a dedicated article on NVIDIA GPU driver installation appears alongside the standard Linux setup guides. That tells you the documentation team is tracking the product roadmap, not just maintaining legacy content.
NovaAI (live chat) is available 24/7 on the portal and website.
Should users be unable to get an answer from NovaAI, users can choose to route to our human assistance (which is the Support team).
As for billing and general enquiries (are on stand-by basis), should there be an urgent case, staff on stand-by will assist as well.
What Makes IP ServerOne Different
Plenty of APAC providers offer competitive cloud specs. What IP ServerOne brings that most don’t is the combination of longevity, compliance depth, and infrastructure that behaves the same wherever you deploy.
1. Twenty-Three Years of Operating History
IP ServerOne has been running managed cloud infrastructure in Malaysia since 2003. Most cloud providers you’re comparing them against didn’t exist in 2003. That history matters when you’re putting production workloads on a platform and need to know the provider will be there when something goes wrong.
2. Four Active Compliance Certifications
ISO 27001 and ISO 27017 for the Malaysian data center. PCI-DSS across Malaysia, Singapore, and Hong Kong. SOC 2 Type II for cloud services and Malaysian data center infrastructure. Malaysia Digital Status accreditation.
For businesses in finance, healthcare, or e-commerce handling card data, these are prerequisites. IP ServerOne has them, and they’re verifiable. That’s meaningful in a region where GDPR-equivalent data sovereignty requirements are tightening year by year.
3. Data Sovereignty Without Compromise
Every server IP ServerOne runs is in Asia. Malaysia, Singapore, or Hong Kong. Your data doesn’t route through a US or European provider as an intermediate. For businesses operating under local data residency requirements, that’s a straightforward win that removes a category of compliance risk entirely.
4. GPU Infrastructure at APAC Pricing
The NovaGPU line runs from RTX 3090 at $303/month up through dual H200 NVL at over $6,300/month, with single, dual, and quad GPU configurations available. The RTX 5090 just landed on the platform. For small to mid-sized AI and ML teams in Southeast Asia that don’t want hyperscaler pricing or hyperscaler complexity, this fills a real gap in the market.
5. Backups Included, Not Sold Separately
Every NovaCloud plan includes tiered automatic snapshots. Hourly for the last 15 hours, daily for the last 7 days, and weekly for the last 4 weeks. That’s not an add-on you pay extra for. It’s part of what a plan costs. For providers that charge separately for backup storage, that inclusion changes the real price comparison significantly.
Who Benefits Most From IP ServerOne
Ideal For
- Running web apps, databases, or APIs in Southeast Asia and need to keep data within the region
- Operating in a regulated industry where ISO 27001, PCI-DSS, or SOC 2 Type II are requirements rather than nice-to-haves
- Building or running AI and ML workloads in APAC and need dedicated GPU access without hyperscaler pricing
- Evaluating providers on long-term stability rather than the lowest headline price this month
- Running Windows Server workloads and need full RDP access, admin rights, and no software restrictions in an APAC data center
Not Ideal For
- Based outside Asia. IP ServerOne’s infrastructure is APAC only. North America or Europe-based users should look at Hetzner, OVHcloud, or DigitalOcean instead
- Expecting a money-back guarantee. IP ServerOne operates a non-refund policy. Start on hourly NovaCloud billing to test the platform before you commit to a longer term
- Relying heavily on self-service documentation. The knowledge base is solid but not as comprehensive as providers like DigitalOcean or Hetzner who’ve built extensive public doc libraries. If you prefer searching docs to opening tickets, factor that in
Why This Earned the Award
HostAdvice’s Top 25 Cloud Hosting recognition goes to providers that hold up under real conditions, not providers that look good in a single-region benchmark screenshot.
IP ServerOne held up across four Linux servers, three regions, and every stressor we ran. Sequential disk reads came back at 238 to 239 MiB/s whether the server was in Malaysia, Singapore, or Hong Kong. CPU scaling stayed linear from 2 cores to 8. Zero stress test failures across every server and every stressor. The Hong Kong region added a cross-border finding that most providers cannot match: 32.3ms to mainland China with less than 0.1ms of variance across the run.
The 90-day uptime hit 99.97% with no unplanned outages. The support team responded in 8 minutes, asked the right clarifying question before answering, and the knowledge base had a GPU driver installation article sitting alongside the standard Linux guides.
Twenty-three years of operating history does not guarantee any of that. But it is consistent with a provider that has had long enough to get the infrastructure right and enough reason to keep it that way.
