Most people searching for a vultr benchmark want one of two answers: whether Vultr's VPS performance justifies the $5–6 monthly spend, or whether its GPU cloud can handle serious inference workloads. Both questions now have hard data behind them. Across a six-month independent test of all nine US datacenters, Vultr's entry Cloud Compute averaged 4,100 on single-thread CPU benchmarks, fourth place among 13 providers, with disk IOPS ranging from 44,500 to 54,200 depending on location. On the GPU side, Vultr's own published runs on NVIDIA HGX B200 hardware pushed 18,829 output tokens per second on the fastest model tested.
Details matter more than the headline, though. Performance varies more between Vultr's own datacenters than between Vultr and its closest rivals, and the inference benchmark contains a nuance most summaries leave out: goodput versus raw throughput. Both are covered below, along with the plans and pricing those numbers were measured on.
If you just want the short version:
General VPS use: the $5–6/mo Cloud Compute plans rank in the top four of a 13-provider CPU comparison, with sub-millisecond intra-datacenter latency
Datacenter choice matters more than the provider gap: New Jersey leads on disk IOPS (54,200 read), Honolulu trails on every metric, and the remaining mainland locations sit within about 4% of each other
AI workloads: Vultr's B200 inference results are unusually trustworthy because the company publishes its full methodology and reproduction scripts
Short answer: Vultr's shared Cloud Compute lands in solid top-tier territory. The aggregate CPU score across nine US locations came out at 4,100 sysbench events per second on a single thread, fourth among 13 providers in the comparison group, behind Hostinger (4,400), Hetzner (4,300) and Kamatera (4,250). Every one of Vultr's nine locations scored above that group's average.
Disk and network numbers hold up as well. Random 4K read IOPS ranged from 44,500 in Honolulu to 54,200 in New Jersey, write IOPS from 35,200 to 43,100, and TCP throughput on the $5 tier landed between 890 and 965 Mbps, effectively saturating gigabit-class ports at the mainland locations. Intra-datacenter round-trip latency measured 0.7 to 1.4 ms. None of that is exotic; none of it is a problem either.
The methodology deserves respect, because most benchmark articles fall apart here. Tests ran on Vultr's $5/mo Regular Performance tier (1 vCPU, 1 GB RAM, 25 GB SSD) with stock Ubuntu 24.04, no kernel tuning and no sysctl tweaks, so the numbers reflect what an actual customer receives. Sessions ran three times a week in every datacenter for six months, five runs per session, with medians taken across roughly 78 sessions per location. The BestUSA VPS project behind the dataset counts that as about 390 individual runs per metric per datacenter, which is how you filter out noisy-neighbor flukes.
For scale: the gap between Vultr and DigitalOcean in the same test group was 2.5%, and 5.1% versus Linode. Keep those percentages in mind for the next section.
This is the part most Vultr benchmark writeups skip, and it is the detail most likely to change where you deploy. Vultr's best CPU score, Silicon Valley at 4,240, beats its worst, Honolulu at 3,780, by 11%. That is a wider spread than the one separating Vultr from DigitalOcean (2.5%) or Linode (5.1%) in the same test. Once you care enough to compare providers, choosing the right Vultr location matters at least as much.
Disk performance tells the same story with sharper edges. New Jersey was the only location that consistently cleared 54,000 read IOPS, and its 43,100 write IOPS runs 9% above the network-wide average, consistent with it being Vultr's flagship US site and first in line for hardware refreshes. Honolulu bottoms out at 44,500 read and 35,200 write, an 18% drop from top to bottom. At that point you are comparing hardware generations inside one provider.
Here is the full six-month picture, grouped by where each location lands:
Top tier
Silicon Valley (SJC): 4,240 CPU, 52,100 read / 41,800 write IOPS, 960 Mbps, 0.8 ms latency
New Jersey (EWR): 4,180 CPU, best disk (54,200 / 43,100) and network (965 Mbps) of the group, 0.7 ms latency
Chicago (ORD): 4,160 CPU, 51,400 / 41,200 IOPS, 955 Mbps, 0.8 ms
Mainland middle pack
Dallas (DFW): 4,120 CPU, 50,800 / 40,500 IOPS, 950 Mbps
Los Angeles (LAX): 4,100 CPU, 49,900 / 40,100 IOPS, 948 Mbps
Seattle (SEA): 4,080 CPU, 49,200 / 39,400 IOPS, 945 Mbps
Atlanta (ATL): 4,050 CPU, 48,600 / 39,000 IOPS, 940 Mbps
A step down
Miami (MIA): 3,960 CPU, 46,800 / 37,500 IOPS, 935 Mbps, 1.1 ms
Honolulu (HNL): 3,780 CPU, 44,500 / 35,200 IOPS, 890 Mbps, 1.4 ms
Three practical rules fall out of that data. Exclude Honolulu and mainland CPU variance drops from 11% to under 4%, so you can simply pick the location nearest your users. If your workload is disk-hungry, databases and busy CMS installs included, New Jersey is the standout. If your users are in Hawaii, Honolulu remains the sensible choice anyway, since no other provider in the comparison group operates a Hawaiian datacenter; just calibrate expectations accordingly.
Vultr also runs datacenters across Europe, Asia and Australia. The public per-location dataset cited here covers the nine US sites.
Vultr's GPU cloud comes with something rare at this price point: a published inference cookbook. The test bed was eight NVIDIA HGX B200 GPUs with 179 GB of HBM3e each, running vLLM, fed 2,048 input tokens and 512 output tokens per request, swept across concurrency levels from 1 to 1,024. Every figure ships with the methodology and the scripts needed to reproduce it, which puts it a step above typical vendor slides.
Peak output throughput across the six models tested:
Nemotron Nano 30B (FP8, 2 GPUs): 18,829 tok/s
Nemotron Nano 30B (NVFP4, 1 GPU): 15,575 tok/s
MiniMax M2.5 229B (4 GPUs): 8,838 tok/s
DeepSeek V3.2 685B (8 GPUs): 4,370 tok/s
Nemotron Super 49B (1 GPU): 3,816 tok/s
GLM-5 744B (8 GPUs): 2,132 tok/s
One honesty note from the same data: Nemotron Super 49B, running on a single GPU, oscillates at some concurrency levels and drops to roughly 1,587 tok/s (about half its peak) at c=256, a batch-scheduling quirk of the single-card setup. Peak throughput holds steady at c=64–128 and c=512.
Latency at 32 concurrent requests, roughly a realistic interactive load, matters more for most applications than peak tok/s. MiniMax M2.5 posted the fastest time to first token at 91 ms despite carrying 229 billion parameters, likely because four GPUs share the prefill work. Nemotron Super 49B followed at 172 ms, and Nemotron Nano 30B FP8 at 206 ms with under 8 ms per output token. The two largest models started slower, 931 ms for DeepSeek V3.2 and 1,341 ms for GLM-5, the latter paying for the attention overhead of its DSA design. Every model kept 99th-percentile inter-token latency under 41 ms, comfortably inside real-time streaming territory.
Raw throughput flatters every GPU deployment, because chasing peak tok/s wrecks latency for everyone queued behind it. Vultr's cookbook therefore also measures goodput, the rate of requests that actually meet defined latency targets. On Nemotron Nano 30B FP8, with an SLO of first token under 500 ms and under 50 ms per output token, goodput peaked at 13.29 req/s with 64 concurrent requests, producing about 8,027 output tok/s. Push to concurrency 512 and raw output climbs to roughly 18,535 tok/s, but only about 6% of requests meet both targets at that load.
Practical takeaway: size your concurrency to your latency budget. For an interactive chat product, the sweet spot in this dataset sits near 64 concurrent users per two-GPU deployment, a long way from the 512 where the raw tok/s figure peaks.
The most cost-efficient configuration in the entire test turned out to be a single B200. Nemotron Nano 30B in NVFP4 quantization delivered 15,575 tok/s on one GPU, or 15,575 tok/s per GPU, against 9,415 per GPU for the FP8 version spread across two cards. That is 1.65 times the per-GPU efficiency, and 58 times more per GPU than GLM-5. The model's Mamba hybrid architecture drives the gap: its state-space layers carry no KV cache overhead and process tokens in linear time, keeping memory bandwidth saturated.
AMD hardware appears in the same cookbook too. Runs on eight AMD Instinct MI325X GPUs, 256 GB of HBM3e per card and two terabytes of VRAM total, saw Qwen3-VL 235B peak near 11,218 tok/s. Having both ecosystems under one account makes side-by-side testing straightforward.
👉 See Vultr's cloud GPU and compute plans
The hardware behind these numbers starts at $5 per month, and Vultr bills its compute and GPU lines by the hour, so the entry barrier is close to zero. Here are the plan families currently listed on the pricing page, with entry specs.
Cloud Compute — Regular Performance: $5/mo
Entry size: 1 vCPU, 1 GB RAM, 25 GB SSD, with larger sizes in the same line
The exact tier the six-month datacenter benchmark tested
Fits light websites, dev boxes and small utilities
👉 Check Regular Performance current pricing
Cloud Compute — High Performance: $6/mo
1 vCPU, 1 GB RAM, 25 GB NVMe SSD at the entry size
NVMe storage on AMD EPYC nodes; the better default for production apps
Same entry price as High Frequency, aimed at steadier workloads
👉 Check High Performance current pricing
Cloud Compute — High Frequency: $6/mo
1 vCPU, 1 GB RAM, 32 GB NVMe SSD at the entry size
Higher-clocked cores for latency-sensitive small applications
Same entry price as High Performance
👉 Check High Frequency current pricing
Optimized Cloud Compute: from $28/mo
Dedicated vCPUs only, so no shared-core variance between neighbors
Built for steady, CPU-bound production loads
Larger sizes available within the line
👉 Check Optimized Cloud Compute pricing
Bare Metal: from $120/mo
Single-tenant dedicated servers with full hardware control
The option when virtualization overhead is unacceptable
Entry configurations start at the $120 tier
👉 Check Bare Metal current pricing
Cloud GPU: hourly billing
NVIDIA HGX B200 (the hardware behind the inference numbers above) and AMD Instinct MI325X available
Pay by the hour, which suits benchmark runs and bursty inference work
Destroy the instance when the job finishes and billing stops
👉 Check Cloud GPU current pricing
Each line scales up through bigger sizes, so treat these figures as entry points and check the full menu for the size you need. Hourly billing caps at the monthly rate: a $5 instance works out to about $0.007 per hour, and it only costs money while it exists.
The cheapest way to settle the question is to replicate the tests yourself, and Vultr's billing makes that almost free. New accounts receive $250 in credits to use during the first 30 days, and hourly billing means a test instance costs cents per session at the low end. Deploy, measure, destroy, repeat in another datacenter if you like.
A workable process, matching what the independent study did:
Deploy a stock Ubuntu 24.04 LTS instance in your candidate datacenter and leave it untuned, or your numbers stop being comparable
Run sysbench on a single thread for 60 seconds to get a CPU score you can line up against the 4,100 reference point
Measure disk with fio using 4K random read and write patterns at queue depth 32
Test network throughput with iperf3 using four parallel streams
Repeat across a few different days if the decision matters; medians beat one-off runs
Destroy the instance when you finish
For inference testing, Vultr publishes the serving presets and benchmark scripts behind its own cookbook numbers, including a full pipeline for downloading a model, serving it and sweeping concurrency levels. Reproducing the B200 results on your own account is mostly copy-and-paste work.
👉 Create a Vultr account and start with the signup credit
For the workloads most people run on a VPS, yes: fourth among 13 providers on CPU, above-average disk and network at all nine tested locations, and sub-millisecond latency inside each datacenter. Where consistency matters more than peak speed, the Optimized line's dedicated vCPUs remove shared-core variance.
Silicon Valley posts the best CPU score at 4,240. New Jersey leads on everything else, with 54,200 read IOPS, 965 Mbps and 0.7 ms latency. The pragmatic answer: skip Honolulu unless your users are in Hawaii, then pick the mainland location closest to them, because the rest sit within about 4% of each other.
Close to nothing. Hourly billing plus the $250 signup credit, valid for the first 30 days, means a typical testing session costs cents, provided you destroy instances when you finish.
There is no conventional time-limited demo, but the signup credit plays the same role: new accounts get $250 in credits valid for their first 30 days, enough to run every shared compute plan and a fair chunk of GPU time without paying up front.
Hosting a site or two on a budget: Cloud Compute Regular Performance at $5/mo or High Performance at $6/mo, in the mainland datacenter nearest your users
Disk-heavy workloads like databases: the same plans, deployed in New Jersey
Steady CPU-bound loads: Optimized Cloud Compute from $28/mo with dedicated vCPUs
LLM inference: Cloud GPU by the hour, starting from the cookbook's published numbers and watching goodput, with raw tok/s as the secondary metric
One fair caveat from the comparison data: if single-thread CPU is your only criterion, Hostinger, Hetzner and Kamatera all scored higher than Vultr in the same test, by margins under 8%. That gap is smaller than the spread between Vultr's own datacenters, which is why location usually deserves more attention than the provider ranking.
Before ordering, check the size you actually need, since every plan family extends well above the entry specs listed here, and confirm current entry pricing on the pricing page because shared tiers do get revised occasionally.
👉 Pick your datacenter and deploy on Vultr