GPU Mart is a GPU cloud hosting brand under Database Mart LLC, a Texas-based infrastructure provider that has been operating since 2005. It specializes almost entirely in GPU rental, with a catalog that runs from entry-level cards for light workloads up to current-generation enterprise accelerators.
Most of what GPU Mart promises on its plan pages held up under real testing, though a few gaps only show up once you actually order, submit a support ticket, and push the hardware under load. Here is the full breakdown.
To evaluate GPU Mart, I applied our hosting review methodology, the same standardized approach we use across all hosting reviews so scores stay consistent, fair, and grounded in real hands-on testing.
Here is how GPU Mart scored across every key parameter.
| Parameter | Score | Why This Score |
|---|---|---|
| Prices | 8.9/10 | No standard refund guarantee, but discounts up to 20 percent on longer billing cycles and broad payment method support. |
| Features | 9.6/10 | A wide catalog for a GPU-focused host, spanning Kubernetes, HA database, load balancing, VPN, and pre-installed AI software stacks. |
| Performance | 8.5/10 | Consistent network results and strong random 4K IOPS, though the CPU generation and memory bandwidth trail newer platforms. |
| Ease of Use | 9.1/10 | Configuration survives account sign-up, and registration is fast, though the billing form appears late in checkout. |
| Support | 9.0/10 | Ticket support answered a technical question in nine minutes, but live chat quality dropped once handed to a human agent. |
| Overall | 9.0/10 | A capable GPU host with real strengths in ordering flow and ticket support, held back by a few specific rough edges. |
GPU Mart’s catalog is built entirely around GPU rental, split across two product types:
Both product types are available across GPU Mart’s full hardware range, from budget-friendly cards like the GTX 1650 and RTX 2060, through mid-tier options like the RTX A4000 and A5000, up to enterprise hardware including the H100, A100, and current-generation Blackwell cards.
See the pricing widget below for the full breakdown across GPU models and billing cycles.
| Plan Name | Space | CPU | RAM | OS | Price | |
|---|---|---|---|---|---|---|
| Express GPU VPS - 2GB | 120 GB | 8 cores | 16 GB | CA$29.15 | Details | |
| Basic GPU VPS - RTX 5060 | 240 GB | 16 cores | 28 GB | CA$117.98 | Details | |
| Professional GPU VPS - RTX Pro 2000 | 240 GB | 16 cores | 28 GB | CA$206.82 | Details | |
| Professional GPU VPS - RTX A4000 | 320 GB | 24 cores | 28 GB | CA$165.18 | Details | |
| Advanced GPU VPS - RTX Pro 4000 | 320 GB | 24 cores | 56 GB | CA$262.34 | Details | |
| Advanced GPU VPS - RTX Pro 5000 | 320 GB | 24 cores | 56 GB | CA$498.31 | Details | |
| Advanced GPU VPS - RTX 5090 | 400 GB | 32 cores | 84 GB | CA$403.92 | Details | |
| Enterprise GPU VPS - RTX Pro 6000 | 400 GB | 32 cores | 84 GB | CA$900.84 | Details |
A few things worth knowing before you order:
If you are testing GPU Mart for the first time, starting on a shorter billing cycle or using the free trial is the safer approach given the non-refundable pricing structure.
The instance I benchmarked was the Professional GPU VPS – RTX A4000, GPU Mart’s entry-level dedicated GPU tier.
Here is what I was working with:
Before getting into the numbers, it matters to place this plan correctly inside GPU Mart’s full catalog. The RTX A4000 sits toward the lower-middle of their lineup. Below it, GPU Mart offers budget cards like the GTX 1650 and RTX 2060 for lighter workloads.
Above it sits a long list of more powerful hardware, including the RTX A5000, A6000, RTX 4090, and at the top end, current-generation Blackwell cards and enterprise accelerators like the H100 and A100. What follows reflects how this specific tier performs. It’s not a ceiling on what GPU Mart can deliver, and it’s not the cheapest thing they offer either.
It is also worth being upfront about what this benchmark suite measures. The tests here cover CPU, memory, disk, and network performance, the infrastructure surrounding the GPU rather than the GPU itself.
A dedicated GPU compute benchmark is a separate test, and I am not folding GPU throughput numbers into this section. What these results tell you is how well-provisioned the rest of the instance is to support GPU workloads, since a fast card sitting behind a slow disk or a congested network still bottlenecks in practice.
I ran the following:
Single-thread: 398.71 events per second, average latency 2.51ms, 95th percentile 3.02ms
Multi-thread (24 threads): 8,465.70 events per second, average latency 2.83ms, thread fairness standard deviation 102.43
The single-thread number is on the low side compared to more modern hardware, and the CPU model tells you why. The Xeon E3-12xx v2 is an Ivy Bridge chip, a generation that dates back to 2012 and 2013.
That does not mean the instance is underpowered for its purpose. GPU-focused plans are rarely sold on raw CPU throughput, and 24 cores gives you plenty of parallel capacity to feed data to the GPU, run preprocessing, or handle multiple concurrent jobs.
The multi-thread scaling is the more useful number here. Going from one thread to 24 delivered a 21.2x increase in throughput, which works out to roughly 88 percent scaling efficiency. That is a reasonable result for a shared virtualized environment, though it falls short of the near-perfect linear scaling I have seen on newer EPYC-based platforms.
The thread fairness standard deviation of 102 on an average of 3,529 events per thread is around 3 percent variance, which tells me the workload was distributed fairly evenly across cores rather than a few cores doing most of the work while others idled.
Sequential Write: 4,837.50 MiB/sec
Sequential Read: 5,219.28 MiB/sec
Both numbers land in a moderate range. As a comparison point, GPU and cloud instances built on current-generation AMD EPYC hardware often clear 6,500 to 7,500 MiB/sec on the same test. This instance comes in noticeably below that band on both write and read.
For most workloads, memory at this speed is not going to be the limiting factor, since GPU training and inference jobs are usually bottlenecked by the GPU itself or by disk and network I/O feeding it. But if you are running memory-heavy preprocessing pipelines or large in-memory datasets, this is a number to keep in mind rather than assume away.
Sequential Write: 442 MiB/s (463 MB/s), IOPS 441
Sequential Read: 1,039 MiB/s (1,089 MB/s), IOPS 1,038
Random 4K mixed read/write: Read ~22,428 IOPS average, Write ~22,475 IOPS average, throughput around 58.5 MiB/s each direction
The gap between sequential write and sequential read is the standout detail in this section. Read speed came in at more than double the write speed, 1,039 MiB/s against 442 MiB/s.
That kind of asymmetry usually points to a storage backend optimized more heavily for read throughput than write throughput, which is common on SSD-backed cloud storage but still worth planning around if your workload involves writing large model checkpoints or datasets to disk repeatedly.
The random 4K result is where this instance performs well. Both read and write IOPS landed above 22,000, which is a strong number for a general-purpose cloud VPS and holds up well against other GPU and cloud instances I have benchmarked. Databases, small file operations, and any workload with unpredictable access patterns will benefit from this.
Run 1: Download 287.37 Mbps, Upload 296.17 Mbps, Idle latency 0.91ms, 0% packet loss
Run 2: Download 286.16 Mbps, Upload 296.59 Mbps, Idle latency 1.07ms, 0% packet loss
Both runs landed within about 1.2 Mbps of each other on download and under 0.5 Mbps on upload, which tells me these results are consistent rather than a one-off. Zero packet loss on both runs is a clean result. Idle latency staying under 1.1ms across both tests is excellent for general connectivity.
One thing worth flagging: the second run printed a “Network unreachable” error at the start of the test before the results populated. Despite that, the test completed and returned a full set of numbers nearly identical to the first run.
This looks like a transient hiccup during test initialization rather than a real connectivity problem, since a genuine network issue would have shown up as inconsistency or packet loss in the final numbers, and it did not. Still, if you are running this test yourself and see the same error, I would treat the result with a bit of caution and run it a third time to confirm.
CPU Stress (180s, 24 workers): 3,786,032 bogo ops, 21,032.57 bogo ops/s, completed successfully
Memory Stress (180s, 2 workers): 4,629,010 bogo ops, 25,685.51 bogo ops/s, completed successfully
Disk I/O Stress (180s, 2 workers): 1,392,640 bogo ops, 7,557.86 bogo ops/s, completed successfully
All three stress phases ran to completion without errors or aborted stressors. This is the test that matters most for anyone planning to run sustained training jobs or long-running inference workloads, since it simulates continuous load rather than a quick burst.
The instance held up cleanly across CPU, memory, and disk simultaneously stressed for three minutes each, which is a reasonable proxy for how it will behave under a real production workload.
The Professional GPU VPS – RTX A4000 performs the way you would expect from GPU Mart’s entry-level dedicated GPU tier: solid rather than exceptional on the surrounding infrastructure, with a real strength on random disk I/O and network consistency.
The CPU is built on older Ivy Bridge silicon, which shows up in the single-thread number, but multi-thread scaling and the clean stress test results suggest the 24 cores are usable rather than a marketing figure.
Memory bandwidth trails what I have seen on newer EPYC-based platforms, and the disk shows a real asymmetry between read and write speed that write-heavy workloads should plan around. None of these are dealbreakers for a plan at this tier, but they are the details that separate this instance from GPU Mart’s higher tiers, and knowing that going in will help you set the right expectations for what this specific plan can do.
GPU Mart operates across two separate interfaces, and that is the first thing to understand before you start an order.
The product catalog, plan pages, and GPU selection all live on the GPU Mart website. Once you click through to purchase, everything shifts to a separate platform called Database Mart, which handles account management, billing, and server control.
Most hosting providers keep this under one roof, so I went in curious whether the split would create friction or hold together across the full journey from first click to active server.
Before configuring anything, I spent time on the GPU Mart website getting a feel for the product range.
The GPU Rental menu in the top navigation opens a three-column dropdown that groups the catalog by generation: Blackwell for current-gen hardware, Previous Gen GPUs for established cards like the H100, A100, and RTX A6000, and an Entry GPUs column for lower-cost options.
Each listing pairs the model name with a short description of what it is suited for. That detail matters more than it might seem. Choosing between an RTX A4000 for inference work versus an RTX Pro 6000 for large-scale training runs is not always obvious from specs alone, and the brief labels help you orient quickly without opening five separate plan pages.
I selected the RTX A4000 and landed on its dedicated pricing page. The plan comparison is clean: two tiers displayed side by side, each listing CPU count, RAM, disk, bandwidth allocation, IP, backup frequency, and location upfront.
Four billing cycles sit across the top with the savings percentage shown for each longer commitment, so the tradeoffs are visible at a glance rather than buried in a calculator somewhere.
One thing I appreciated about the ordering flow is the sequence. You configure your server fully before the platform asks you to log in or create an account.
I clicked through to the order page, selected Ubuntu Server 24 LTS as the OS, named the instance, and set the billing cycle, all before any sign-in prompt appeared.
Before moving on, the Operating System & Software tab deserves attention for anyone coming to GPU Mart with an AI or ML workload. Alongside a standard OS selection, it offers a set of pre-installed application stacks:
If you want to land on a server with an inference or training environment already in place, this tab is how you do it. It is a useful shortcut that saves manual setup time and is easy to miss if you stay on the default OS tab.
When I clicked through to checkout, the page displayed a note that I would need to log in to complete the order. This is the handoff point to Database Mart.
The signup form is minimal: email address, password, and a terms checkbox. Google and GitHub sign-in options are available as well, which is a practical choice for developers who would rather not manage another set of credentials.
Alt: Database Mart sign-up page showing email and password fields with Google and GitHub sign-in options and platform stats
Alt: Database Mart registration success screen showing congratulations message after account creation
What happened after completing the signup is the most positive detail in this entire process. I was taken directly back to the configuration page I had been on, with every choice fully intact: the plan, OS, server name, billing cycle, everything.
On several other platforms I have tested, completing a sign-up mid-order sends you back to the start, and you lose the configuration. Database Mart does not do that, and it makes a real difference to the experience when you have spent time making selections.
Clicking Checkout brought me to an order confirmation screen, where the billing information was flagged as incomplete.
This is where a new customer hits their first form modal, entering their name, address, and phone number before payment goes through. It is standard information to collect, but it appears at checkout rather than during account setup, which means it interrupts the payment flow at what feels like an inconvenient moment for a first-time user.
Once the billing form was submitted, the invoice page offered two payment options: credit card or PayPal. No setup fees appeared on the order. I paid, received a confirmation email immediately, and the server appeared in my dashboard right away.
What I thought about the registration process: It is leaner than most GPU hosting providers I have tested. The decision to let you build out a full server configuration before asking for an account is the standout design choice, and it works well because the configuration survives the sign-up round-trip intact.
The friction is limited to two points: billing information appearing at checkout rather than during account creation, and the mid-purchase shift from the GPU Mart website to a separately branded Database Mart portal. Neither breaks the flow, but both create a moment of “is this right?” that a smoother handoff would eliminate.
After payment cleared, the Database Mart dashboard opened on a Guidance screen rather than a blank account page.
That was the right call. It shows a three-step progression: Account Status, Add a Payment Method, and Deploy Your First Server. For anyone landing here with no active services, this tells you exactly where to go next without assuming any prior familiarity with the portal.
Below the guidance steps, the dashboard breaks into two panels. Product Overview presents the six service categories as tiles: Virtual Servers, Dedicated Servers, GPU Servers, Domains, SSL Certificates, and All Products, each with a short description of what it covers.
Pre-Purchase Guidance sits below that with links to promotional offers, a pre-sale FAQ, a server selection tool, the App Hub for one-click deployments, and a free trial application. These are functional entry points rather than decorative ones.
The left sidebar covers the full navigation: Home, My Services, Customer Support, Customer Tools, Billing, Account, All Products, and Guidance. Everything is labeled clearly, and nothing is more than one level deep.
One feature I want to flag at the bottom of the dashboard is a Pre-Sales Inquiry text field. It is an uncommon placement, but the intent makes sense. If you are looking at the product catalog and have a question before committing, you can ask without leaving the page. It is not a replacement for live chat, but it is a practical shortcut that many providers do not think to include at this level of the interface.
The one gap I noticed on first login was that my active server did not appear on the home screen. It lives under My Services rather than being surfaced directly in the dashboard view. A returning customer managing multiple instances will know to look there, but someone who just paid and is wondering where their server went will experience a brief moment of uncertainty before finding it.
What I thought about the dashboard: The Guidance page on first login is the right instinct. Rather than landing on an empty account screen, you get a structured set of next steps that map directly to what a new customer needs to do. The sidebar is clean, the product overview makes sense, and small touches like the Pre-Sales Inquiry field show that some thought went into the post-signup experience.
The one thing I would change is surfacing your active server on the home screen after payment rather than requiring a trip to My Services. It’s a minor gap, but it lands at the worst possible moment for a new customer.
The two-brand setup holds together better than it looks on paper. The main rough edges are both in checkout: billing information collected too late, and a mid-order brand switch that catches new users off guard. Neither is a dealbreaker.
Once inside Database Mart, experienced GPU cloud users will find their footing quickly, and the Guidance page gives newcomers enough of a starting point that they are not left staring at a blank screen.
GPU Mart and Database Mart share support infrastructure, and the channel lineup is broader than what you might expect from a specialist GPU host. Before testing anything directly, I mapped out what was available:
I tested ticket support and live chat directly, using different technical questions across each channel.
To open a ticket, I went to Customer Support in the left sidebar of the Database Mart portal. The page shows your ticket history alongside two buttons in the top right: Online Chat and Submit Ticket.
Clicking Submit Ticket opens a modal with fields for Subject, Department, Service, IP/Domain, CC, and a rich text message body. The department defaults to Support Department. File attachments are supported, which is a useful option if you need to share logs or terminal output alongside your question.
I submitted this question at 2:58 AM on 6 August 2026:
Subject: Bandwidth cap on GPU VPS (ingress vs egress and burst behavior)
“Hi. The RTX A4000 plan lists 300Mbps unmetered bandwidth. Is that limit applied to ingress, egress, or is it a combined bidirectional cap? And if I am pulling a large dataset or model weights into the instance at once, is there any burst headroom above that rate, even temporarily?”
The response came at 3:07 AM, nine minutes later, from Alina in the Support Department.
She answered both parts without needing a follow-up. The 300Mbps cap applies to both inbound and outbound traffic as a single combined limit, not separately per direction. There is no burst headroom above that ceiling.
She also added detail I had not asked for: that actual throughput can vary depending on load across the physical rack, since bandwidth is shared at the infrastructure level. Then she went further and flagged that bandwidth add-on options exist for customers with higher requirements, and suggested scheduling large transfers during off-peak hours as a practical workaround in the meantime.
What I thought about ticket support: The nine-minute turnaround at close to 3 AM is the standout result. A lot of providers claim 24/7 support but deliver response times outside business hours that run into hours, not minutes. Alina addressed both parts of the question clearly, volunteered information about rack-level bandwidth sharing without being asked, and closed with a concrete next step. That combination of speed and depth is not common at this price tier.
Live chat is not built into the Database Mart portal. It lives on the GPU Mart website, where a Chat with Human Support widget appears in the bottom right corner with an Online status indicator.
Clicking it opens a short pre-chat form asking for your name, email, inquiry type, and a description of the issue before connecting you.
I submitted this question at 1:10 PM:
“Hi. I am planning to run a workload that requires CUDA 12.6. Can you confirm whether the Ubuntu 24 GPU image ships with a driver version that supports that, and if the installed driver is older, what is the recommended way to upgrade it without breaking the existing CUDA installation?”
Clarity AI, the platform’s AI assistant, responded within seconds. The answer was specific: the Ubuntu 24 GPU image ships with NVIDIA driver version 535, which supports CUDA 12.x. It provided the exact commands to remove the current driver and reinstall the required version, and linked to three knowledge base articles for reference.
At the bottom of the AI response, three options appeared: End Chat, Submit Ticket, and Chat with Agent.
I clicked Chat with Agent, and Robert connected within one minute.
His response arrived at 1:12 PM: “My recollection is that the CUDA version should be higher than 12.6, but this may depend on whether the GPU supports it. If the GPU supports it, then the CUDA version should be higher than 12.6.”
What I thought about live chat: The AI handled this better than the human agent who followed. Clarity AI named a specific driver version, gave actual commands, and pointed to documentation. Robert’s response leaned on personal recollection rather than confirmed information and repeated the same point twice without resolving the question.
The connection speed was good, which makes the answer quality more noticeable by contrast. If you use live chat, the AI first-response is the more reliable layer for GPU-related questions. For anything that requires a definitive technical answer, the ticket channel is the better route.
The Database Mart knowledge base is hosted on the main website and does not require a login to access. The homepage opens with a search bar and a set of featured articles grouped by OS: Linux and Windows.
Featured pieces include guides on enabling root SSH on Debian and Ubuntu, opening ports in Windows Firewall, detecting and removing XMRig crypto mining malware, running a GPU stress test with GPU Burn, and changing the administrator password on Windows.
Below the featured section, a longer list covers a wider range of topics: LAMP stack installation, DHCP configuration, FTP user management, SSL certificate setup, and more. The search bar at the top is functional and returns relevant results when queried with specific terms.
I opened the XMRig crypto mining malware detection article to check how the content is written.
It’s well put together: the structure moves from introduction to an explanation of why the threat matters, then into a multi-step detection guide with screenshots at each stage, command line examples in formatted code blocks, a tools section, and a conclusion. The article is written for someone who is mid-problem and needs to act, not someone who wants background reading.
What I thought about the knowledge base: The writing quality in the articles I reviewed is solid. The steps are accurate, the code blocks are formatted cleanly, and the screenshots are placed where they add clarity rather than just padding.
My main observation is that the catalog is weighted toward general Windows and Linux server administration rather than GPU-specific topics. For a provider whose primary product is GPU hosting, the knowledge base has a gap: there are no articles on CUDA toolkit management, GPU driver troubleshooting, or workload-specific configuration. That is the content most likely to be searched by the customer base using this service.
Ticket support is where Database Mart’s support operation performs best. A nine-minute response to a technical question in the early hours of the morning, from an agent who answered everything asked and added relevant detail beyond it, is a strong result.
Live chat is faster to start but less consistent once it moves past the AI layer: Clarity AI gave a better first response to my technical question than the human agent who followed.
The knowledge base is well written but covers general server topics more than GPU-specific ones, which is a gap that matters for the audience most likely using this platform. For critical or complex technical questions, the ticket system is the channel to use.
Yes, with some conditions. GPU Mart is a solid choice if you need GPU compute without navigating a hyperscaler’s pricing structure, and my testing backs that up. The ordering flow is smoother than it looks on paper, ticket support answered a real technical question in nine minutes with a complete answer, and the RTX A4000 instance held up cleanly under sustained stress testing.
Where GPU Mart falls short is the refund policy and the inconsistency in live chat once a human agent takes over from the AI.
If either of those matters to you, lean on the one-day free trial before committing, and use the ticket system over live chat for anything technical. For developers and small teams who need GPU access without a long-term commitment or enterprise pricing, GPU Mart is a provider worth putting on the shortlist.
| Plan Name | Space | CPU | RAM | OS | Price | |
|---|---|---|---|---|---|---|
| Express GPU VPS - 2GB | 120 GB | 8 cores | 16 GB | CA$29.15 | Details | |
| Basic GPU VPS - RTX 5060 | 240 GB | 16 cores | 28 GB | CA$117.98 | Details | |
| Professional GPU VPS - RTX A4000 | 320 GB | 24 cores | 28 GB | CA$165.18 | Details | |
| Professional GPU VPS - RTX Pro 2000 | 240 GB | 16 cores | 28 GB | CA$206.82 | Details | |
| Advanced GPU VPS - RTX Pro 4000 | 320 GB | 24 cores | 56 GB | CA$262.34 | Details | |
| Advanced GPU VPS - RTX 5090 | 400 GB | 32 cores | 84 GB | CA$403.92 | Details | |
| Advanced GPU VPS - RTX Pro 5000 | 320 GB | 24 cores | 56 GB | CA$498.31 | Details | |
| Enterprise GPU VPS - RTX Pro 6000 | 400 GB | 32 cores | 84 GB | CA$900.84 | Details |
Yes, for GPU compute in particular. GPU Mart offers a wide range of GPU hardware, from budget cards to enterprise accelerators, backed by responsive ticket support and an ordering process that does not overcomplicate things. The main drawbacks are the lack of a money-back guarantee and inconsistent live chat quality once escalated to a human agent.
No. GPU Mart operates under Database Mart LLC’s standard terms, which state that all fees are non-refundable except at the company’s discretion. If you want to test the platform risk-free, use the one-day free trial before purchasing a plan.
GPU Mart’s catalog spans three tiers: entry-level cards like the GTX 1650 and RTX 2060, previous-generation options including the RTX 4090, A100, and A6000, and current-generation Blackwell hardware such as the RTX Pro 6000. Both GPU VPS and dedicated bare-metal GPU servers are available across most of the lineup.
Yes. GPU Mart offers a one-day free trial on VPS, dedicated, and GPU server products, though valid billing information is required before a trial is approved. Each user can request up to three trial orders.
GPU Mart is the product-facing brand for GPU rental, while Database Mart is the account and billing platform behind it. You browse and configure GPU plans on the GPU Mart website, then complete registration, checkout, and ongoing account management through the Database Mart portal.
Answer a few simple questions and find the perfect solution for you!
Start Hosting Search
