GPU servers · RTX 50 series
RTX 5090 dedicated server with 32 GB for your models and renders
An RTX 5090 dedicated server is a whole machine you rent by the month with NVIDIA’s RTX 5090 inside: 32 GB of GDDR7, passed straight through to your own operating system. Nobody else uses the GPU or the server, the port is unmetered, and there is no contract. If 16 GB is enough for your job, the RTX 5080 goes into the same racks for less.
| Figure | RTX 5090 | RTX 5080 |
|---|---|---|
| GPU memory | 32 GB GDDR7 | 16 GB GDDR7 |
| Memory bus | 512-bit | 256-bit |
| CUDA cores | 21,760 | 10,752 |
| Boost clock | 2.41 GHz | 2.62 GHz |
| AV1 encoders | 3 | 2 |
| Power | 575 W | 360 W |
| GPU interface | PCIe Gen 5 | PCIe Gen 5 |
Own an RTX 50 series machine you are not using?
List it on the Primcast marketplace, pick your own monthly price, and let people who need the compute rent it from you.
00Marketplace
Or rent an RTX 50 series GPU from the person who owns it
Owners put their RTX 50 series machines on our marketplace and set the monthly price themselves. A listing marked as held by our sister company, Server Room, is rented on serverroom.net instead of here.
01What it is for
What you can run on an RTX 5090 AI server
This is NVIDIA’s biggest consumer GPU, in a rack. You get a lot of compute and memory for the money, and the limits that come with consumer hardware.
01AI models
Inference and fine-tuning in 32 GB
A 30-billion-parameter model at 4-bit fits with room to spare, and so does a 13-billion model at 8-bit. The same 13-billion model at full 16-bit does not leave enough room to serve it. The sizing tables below show every case.
02Rendering
Big scenes that stay in GPU memory
GPU renderers are fast as long as the scene fits on the GPU. 32 GB holds far more geometry and textures than the 24 GB GPUs a step down, and a batch of frames can run overnight while your own workstation stays free.
03Video
Three AV1 encoders on one GPU
The RTX 5090 has three hardware video encoders and two decoders, so it can transcode several streams at once. The A100 and H100 accelerators have no encoder at all. For live channels, see our transcoding servers.
04Limits
No error-correcting memory
GeForce GPUs have no ECC and no certified professional driver. That is fine when a bad result means a rerun. For runs that last weeks without checkpoints, or software licensed only on professional GPUs, look at the RTX PRO 6000 with 96 GB and ECC.
02Specifications
NVIDIA RTX 5090 and RTX 5080 specifications
The figures NVIDIA publishes for the two RTX 50 series GPUs we fit. Both use the Blackwell architecture, with fifth-generation Tensor Cores and fourth-generation RT Cores.
RTX 5090
Blackwell
32 GB GDDR7
21,760
512-bit
2.41 GHz
575 W
PCIe Gen 5 x16
3 NVENC (9th generation), 2 NVDEC, AV1
No
RTX 5080
Blackwell
16 GB GDDR7
10,752
256-bit
2.62 GHz
360 W
PCIe Gen 5 x16
2 NVENC (9th generation), 2 NVDEC, AV1
No
03Before you order
The PCIe slot your RTX 5090 gets in each machine
Both GPUs are built for PCIe Gen 5. Every platform we put them in is Gen 3 or Gen 4, so the link between GPU and server runs below what the GPU was designed for. Here is what each machine gives it.
Intel Xeon E5-2600 v3/v4
Intel Xeon Silver / Gold
AMD EPYC
We say this up front because it has caught a customer before: an RTX 5090 on a Gen 3 platform rendered slower than the RTX 4090 it replaced. A slower slot only hurts when your job moves data across it all the time, like a training loop that loads a fresh batch from the host every step. Inference, transcoding and game servers barely touch it once everything is loaded. Not sure which yours is? Tell us on chat what you will run and we will point you to the right machine.
04The smaller one
RTX 5080 server: the same generation with 16 GB
The RTX 5080 has the same Blackwell design, the same generation of AV1 encoders and the same PCIe Gen 5 interface, with about half the CUDA cores and half the memory of the RTX 5090.
When the RTX 5080 is enough
Your model fits in 16 GB with room left over: a 7-billion model at 8-bit, or a 13-billion model at 4-bit. Or the job is video, where its two AV1 encoders do the work, or a game server that never fills the GPU’s memory.
When to pay for the RTX 5090
Your model or scene is bigger than 16 GB, or you want the extra memory for longer contexts and more users at once. Twice the memory on a bus twice as wide is the real difference; clock speeds are close.
05Sizing
What fits in 32 GB and in 16 GB of GPU memory
Weights only. Leave room on top for the context and the runtime.
RTX 5090, 32 GB
| Model | 16-bit | 8-bit | 4-bit |
|---|---|---|---|
| 7-8 billion parameters | 16 GB: fits | 8 GB: fits | 5 GB: fits |
| 13-14 billion parameters | 28 GB: does not fit | 14 GB: fits | 8 GB: fits |
| 30-34 billion parameters | 68 GB: does not fit | 34 GB: does not fit | 19 GB: fits |
| 70 billion parameters | 140 GB: does not fit | 70 GB: does not fit | 38 GB: does not fit |
RTX 5080, 16 GB
| Model | 16-bit | 8-bit | 4-bit |
|---|---|---|---|
| 7-8 billion parameters | 16 GB: does not fit | 8 GB: fits | 5 GB: fits |
| 13-14 billion parameters | 28 GB: does not fit | 14 GB: does not fit | 8 GB: fits |
| 30-34 billion parameters | 68 GB: does not fit | 34 GB: does not fit | 19 GB: does not fit |
| 70 billion parameters | 140 GB: does not fit | 70 GB: does not fit | 38 GB: does not fit |
Power is the other limit. One RTX 5090 draws up to 575 W, which is why a second one in the same machine is a build we quote, not an option at checkout. To start from your model and find the GPU it needs, use the LLM memory sizing page.
06Choosing
RTX 5090 server, RTX 4090D or RTX 3090?
These GPUs go into the same machines, so newer is not automatically better for you. Compare memory and the monthly price for your job.
| GPU | GPU memory | GPU alone | Cheapest complete machine |
|---|---|---|---|
| RTX 3090 | 24 GB GDDR6X | $239 | $309.93 |
| RTX 5080 | 16 GB GDDR7 | $229 | $312.67 |
| RTX 4090D | 24 GB GDDR6X | $299 | $372.47 |
| RTX 5090 | 32 GB GDDR7 | $399 | $472.47 |
The RTX 4090D and the RTX 3090 both have 24 GB. They are PCIe Gen 4 GPUs, so the slot costs them less than it costs the RTX 50 series. If your work fits in 24 GB, read the RTX 4090 dedicated server page and the RTX 30 series servers page before you decide.
07Prices
RTX 5090 and RTX 5080 server prices, with the machine
Your invoice shows the GPU on one line and the server on another. Both come from the same order catalogue the checkout uses, and each price below belongs to the machine next to it.
| GPU | The machine it goes in | GPU alone | Complete, per month |
|---|---|---|---|
| RTX 5080 | Intel Xeon Silver / Gold | $229 | $312.67 Configure this build → |
| RTX 5090 | Intel Xeon E5-2600 v3/v4 | $399 | $472.47 Configure this build → |
| RTX 5090 | Intel Xeon Silver / Gold | $399 | $482.67 Configure this build → |
| RTX 5090 | AMD EPYC | $399 | $616.59 Configure this build → |
Traffic is not metered in either direction and there is no egress fee. Every server comes with an IPv4 address, you pay month to month, and you can leave at the end of any month. To see every GPU we fit, priced the same way, go to all GPU servers with prices.
08Delivery and help
When your RTX 5090 server is ready, and who helps you
Delivery depends on where the GPU is when you order. Ask on chat before you pay and we tell you which of these applies to your build.
- About 30 minutes
Already in a rack
If a racked machine already has the GPU fitted, you get it about 30 minutes after your order is approved.
- 4 to 24 hours
On our shelf
We have the GPU in hand and fit it into the server you chose.
- 5 to 10 working days
Ordered for you
We buy the GPU from a distributor and fit it as soon as it arrives.
Ask anything on our chat, before or after you order. An AI assistant replies in seconds and can check your account and your server, and a person from our team steps in whenever you ask. Machines that are racked and ready now are on instant servers.
09Questions
Questions about renting an RTX 5090 server
01What do I get with an RTX 5090 dedicated server?
A whole physical server with one RTX 5090 inside, rented by the month. The full GPU is passed through to your own operating system, with no hypervisor, no vGPU split and no time-sharing. You install the driver and CUDA version you want, and nothing changes them behind your back.
02Is there an RTX 5090 Server Edition?
No. NVIDIA does not make a server edition of the RTX 5090. Its passively cooled Blackwell GPU for servers is the RTX PRO 6000 Blackwell Server Edition, with 96 GB of memory and ECC. We fit the RTX PRO 6000 too: see the RTX PRO 6000 dedicated server page. The RTX 5090 we rent is NVIDIA’s GeForce part, running in a rack server.
03Will an RTX 5090 run at full speed in your servers?
Its PCIe link will not. The RTX 5090 is a PCIe Gen 5 GPU and our platforms are Gen 3 and Gen 4, so the bus runs at a half or a quarter of its design speed, depending on the machine. That slows training that streams data from the host every step. It makes little difference to inference, transcoding or rendering once the work is loaded into GPU memory. The slot each machine gives is listed on this page.
04How big an AI model can an RTX 5090 server run?
With 32 GB, a 30-billion-parameter model fits at 4-bit and a 13-billion model fits at 8-bit, both with room for the context. A 70-billion model does not fit on one RTX 5090, not even at 4-bit. For bigger models, see the LLM memory sizing page.
05RTX 5090 or RTX 5080: which should I rent?
Rent the RTX 5080 if your model or scene fits in 16 GB with room to spare, or if the job is mostly video. Rent the RTX 5090 when you need more than 16 GB, or want headroom for longer contexts and more users. The price table above shows what each costs with its machine.
06Can I get an RTX 5070 or RTX 5060?
No. The RTX 5090 and the RTX 5080 are the only RTX 50 series GPUs in our order catalogue, and the configurator offers no others. For a smaller, cheaper GPU, compare the rest on all GPU servers with prices.
07Can I put two RTX 5090 GPUs in one server?
Yes, as a build we quote rather than a checkout option. Each one draws up to 575 W, and GeForce RTX 50 GPUs have no NVLink, so two of them are two separate 32 GB pools. Tell us what you plan to run, or see multi-GPU servers.
08How soon can I have an RTX 5090 server?
About 30 minutes after approval if a racked machine already has the GPU, 4 to 24 hours if we have the GPU on the shelf, and 5 to 10 working days if we have to buy it in. Ask on chat before you pay and we tell you which one applies.
Choose your RTX 5090 build and see the monthly total
Every build above is priced with its server. Pick one, set it up the way you want, and the order shows your full monthly price before you pay.