GPU servers · RTX 40 series
RTX 4090 dedicated server and the GPU in it is the RTX 4090D
Our RTX 4090 dedicated server carries the RTX 4090D, the edition NVIDIA made for China: the same 24 GB of GDDR6X and the same memory speed, with about 11% fewer CUDA cores. You rent the whole machine by the month, the full GPU is passed through to your own system, and the port is unmetered.
| Figure | RTX 4090 | RTX 4090D |
|---|---|---|
| CUDA cores | 16,384 | 14,592 |
| Tensor cores | 512 | 456 |
| Memory | 24 GB GDDR6X | 24 GB GDDR6X |
| Bandwidth | 1,008 GB/s | 1,008 GB/s |
| Boost clock | 2.52 GHz | 2.52 GHz |
| AV1 encoders | 2 | 2 |
| Board power | 450 W | 425 W |
Have an RTX 40 series GPU sitting idle?
Put the machine on the Primcast marketplace, set your own monthly price, and rent it to people who need the compute.
00Marketplace
Or rent an RTX 40 series GPU from its owner
Owners list these machines on our marketplace and name their own price. A listing held by our sister company, Server Room, is rented on serverroom.net, not here.
01RTX 4090D vs RTX 4090
What the D in RTX 4090D means for your work
In late 2023 a United States export rule put the RTX 4090 over the line for sale in China. NVIDIA answered with the RTX 4090D, released on 28 December 2023, with 14,592 CUDA cores where the full GPU has 16,384.
What stays the same
The memory is untouched: 24 GB of GDDR6X on a 384-bit bus at 1,008 GB/s. So is the 2.52 GHz boost clock and the pair of AV1 encoders. Anything limited by memory size or memory speed behaves exactly as it would on a full 4090: a model that fits in 24 GB still fits, and it reads just as fast.
Drivers and software do not notice. It is an Ada Lovelace GeForce GPU, it takes the normal GeForce driver, and every CUDA release, framework and renderer that runs on a 4090 runs on it.
What you give up
About 11% of the shaders and tensor cores. NVIDIA put the real-world gap at about 5% in games and creative apps, because clocks and memory carry much of the load. On work that keeps every shader busy, such as a long render or a training run, plan for up to a tenth less than a full 4090.
We tell you this before you order rather than printing “RTX 4090” and letting you find out later. If your job truly needs the full GPU, ask on chat and we will tell you whether we can source one and how long it takes.
02What it is for
What a 4090 GPU server does well, and where it stops
This is a consumer GPU in a rack. That is what keeps the price down, and it is also where its limits come from.
01Compute per dollar
A lot of compute for the money
GeForce silicon is one of the cheapest ways to rent this much compute. It suits work where a bad result means a rerun, not a disaster: batch rendering, game servers, fine-tuning small models, inference you can repeat.
02Video
It encodes video, the accelerators do not
Two hardware encoders with AV1, plus a decoder, so it can transcode all day. The A100 and H100 further up our catalogue have no video encoder at all. For anything with video in it, this GPU is the right choice, not the compromise. See our transcoding servers for live channel work.
03Memory
The limit is 24 GB
A 13-billion-parameter model fits at 8-bit, the same model at 16-bit does not, and a 70-billion model does not fit at any precision. The sizing table further down is exact about it.
04Long runs
No error-correcting memory
There is no ECC and no certified professional driver. For a run that lasts weeks with no checkpoints, or software whose licence asks for a professional GPU, look at the L40S with 48 GB and ECC instead.
03Specifications
NVIDIA RTX 4090D specifications
The figures for the GPU we fit, from NVIDIA’s RTX 40 series documentation and NVIDIA’s encode and decode support matrix.
RTX 4090D
Ada Lovelace
24 GB GDDR6X
14,592
1,008 GB/s
425 W
PCIe Gen 4 x16
2 NVENC, 1 NVDEC, AV1
No
04Choosing
RTX 4090 server, RTX 3090 or RTX 5090?
All three are on our shelf today and go into the same machines. So the question is not which one is newest, but which one is worth the difference for your job.
| GPU | GPU memory | The GPU | Cheapest complete machine |
|---|---|---|---|
| RTX 3090 | 24 GB GDDR6X | $239 | $309.93 |
| RTX 4090D | 24 GB GDDR6X | $299 | $372.47 |
| RTX 5090 | 32 GB GDDR7 | $399 | $472.47 |
The RTX 3090 has the same 24 GB and is the older, slower way to get it. If memory size is what you need and speed matters less, read the RTX 30 series servers page first. The RTX 5090 moves up to 32 GB of GDDR7. It is a PCIe Gen 5 GPU going into Gen 3 and Gen 4 platforms, so read the next section and the RTX 5090 servers page before you assume newer means faster here.
05The machine
Which machine the RTX 4090D goes in, and its PCIe link
The 4090D is a PCIe Gen 4 GPU. Each platform below shows the slot it gets and what that means in practice.
Intel Xeon E5-2600 v3/v4
Intel Xeon Silver / Gold
AMD EPYC
We have seen a customer move to a newer GPU on an older platform and render slower than before. Once the scene or the weights sit in GPU memory, the slot is quiet and its generation stops mattering. It matters when your job pushes data across the slot on every step. Tell us what you will run and we will point you to the right machine.
06Sizing
What fits in 24 GB of GPU memory
Weights only. Leave room for the context and the runtime on top.
| 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, tight |
| 70 billion parameters | 140 GB, does not fit | 70 GB, does not fit | 38 GB, does not fit |
The other limit is power. One 4090D draws up to 425 W, which is why more than one in a machine is a build we quote, not a box you tick. To work the question the other way round, from your model to the GPU it needs, use the LLM memory sizing page.
07Prices
RTX 4090D server prices, with the machine
The GPU is one line on your invoice and the server is another. Both come from the order catalogue the checkout uses, so every price below belongs to the machine beside it.
| GPU | The machine it goes in | The GPU | Complete, per month |
|---|---|---|---|
| RTX 4090D | Intel Xeon E5-2600 v3/v4 | $299 | $372.47 Configure this build → |
| RTX 4090D | Intel Xeon Silver / Gold | $299 | $382.67 Configure this build → |
| RTX 4090D | AMD EPYC | $299 | $516.59 Configure this build → |
The port is unmetered in both directions, with no egress charge. An IPv4 address comes with the server, there is no contract, and you pay one month at a time. To compare every GPU we fit, priced the same way, see all GPU servers with prices.
08Delivery and help
How soon you get it, and who answers
It depends on where the GPU is when you order. You are told which of these applies before you pay.
- About 30 minutes
Already racked
A machine with the GPU fitted is standing in a rack. It is yours about 30 minutes after the order is approved.
- 4 to 24 hours
On our shelf
We fit a GPU we already hold into the server you picked.
- 5 to 10 working days
Bought in
We order the GPU from a distributor for you, then fit it to your server.
Questions before or after you order go to our chat. An AI assistant answers in seconds and can look up your account and your server, and a person from our team joins whenever you ask. Machines racked and ready right now are on instant servers.
09Questions
Questions about renting an RTX 4090 server
01Is the RTX 4090D the same as an RTX 4090?
Not quite. It is the same AD102 chip with the same 24 GB of GDDR6X, the same 1,008 GB/s and the same boost clock, but with 14,592 CUDA cores instead of 16,384. NVIDIA made it for China after a United States export rule covered the full 4090. It is a genuine NVIDIA GPU, not a rebadge.
02How much slower is the RTX 4090D?
NVIDIA put it at about 5% in games and creative apps. On work that keeps every shader busy, the gap can reach the 11% difference in core count. Anything limited by memory size or memory speed runs the same as on a full 4090.
03Can I get a full RTX 4090 instead?
Not from this page. What we fit is the 4090D, and we say so up front. If your work truly needs the extra shaders, tell us on chat what you will run and we will say whether we can source a full 4090 and how long it would take.
04Does the 4090D work with CUDA, PyTorch and the normal drivers?
Yes. It reports as an Ada Lovelace GeForce GPU and takes the standard GeForce driver. Every CUDA release, framework build and renderer that runs on a 4090 runs on it. The difference is a core count, not an architecture.
05Is the GPU shared with anyone else?
No. You get the full GPU on a machine nobody else uses, passed straight through to your own operating system. There is no hypervisor, no vGPU split and no time-sharing. You install and pin your own driver and CUDA version, and nothing upgrades behind your back.
06Should I rent an RTX 4090D or an RTX 3090 for 24 GB?
If memory size is all you need, the RTX 3090 is often the better buy: same 24 GB, older and usually cheaper per month, as the comparison table shows. The 4090D is clearly faster and adds AV1 encoding, which the 3090 lacks. Pick it when your runs are back to back or you work with video.
07Can I put two RTX 4090D GPUs in one server?
Yes, as a quoted build rather than a checkout option. Each one draws up to 425 W and there is no NVLink, so two of them are two separate 24 GB pools. Tell us what you plan to run and we come back with the machine, or see multi-GPU servers.
08How soon can I have an RTX 4090 dedicated server?
About 30 minutes after approval if a racked machine already carries the GPU, 4 to 24 hours if we hold one on the shelf, and 5 to 10 working days if we buy it in. You are told which applies before you pay. Racked machines are listed on instant servers.
Pick your RTX 4090D build and see the full price
Every build is priced above with its server. Choose one, configure it, and the order shows your monthly total before you pay.