GPU dedicated servers · NVIDIA Ampere · by the month
Our RTX 3090, 3080, 3070 GPU servers are bare-metal machines rented by the month, with the full GPU passed through to your own operating system. Choose one of five RTX 30 GPUs with 8 to 24 GB of memory, on an Intel Xeon or AMD EPYC machine, from $209.93 a month complete.
The price on each GPU below is the complete server: processor, memory, disks, the port and one IPv4 address, plus the GPU's own monthly line. Pick the GPU by the memory your work needs, then change anything else in the configurator.
The five GPUs
All five share the Ampere architecture, the driver and the feature set: 3rd-generation Tensor cores, 2nd-generation RT cores, PCIe 4.0. What separates them is memory size and memory bandwidth. Start from the memory you need and stop at the first GPU that holds it.
8 GB GDDR6
$209.93/mo, complete
GPU $139/mo, plus the machine:
Intel Xeon E5-2630L, 6 cores, 2.00 GHz · 32 GB DDR3 · 2 × SATA 500 GB (RAID 1) · 500 Mbps
8 GB GDDR6X
$222.47/mo, complete
GPU $149/mo, plus the machine:
Intel Xeon E5-2620 v4, 8 cores, 2.10 GHz · 32 GB DDR4 · 2 × SATA 500 GB (RAID 1) · 1 Gbps
10 GB GDDR6X
$229.93/mo, complete
GPU $159/mo, plus the machine:
Intel Xeon E5-2630L, 6 cores, 2.00 GHz · 32 GB DDR3 · 2 × SATA 500 GB (RAID 1) · 500 Mbps
12 GB GDDR6X
$272.47/mo, complete
GPU $199/mo, plus the machine:
Intel Xeon E5-2620 v4, 8 cores, 2.10 GHz · 32 GB DDR4 · 2 × SATA 500 GB (RAID 1) · 1 Gbps
24 GB GDDR6X
$309.93/mo, complete
GPU $239/mo, plus the machine:
Intel Xeon E5-2630L, 6 cores, 2.00 GHz · 32 GB DDR3 · 2 × SATA 500 GB (RAID 1) · 500 Mbps
Each price is the cheapest machine that takes that GPU, as the configurator opens it. The button opens that machine with the GPU already selected; other processors, more memory, larger disks and a faster port are priced there. Bucharest builds are billed in euros, and the configurator shows that price when you choose the city.
Sizing
For inference the first question is memory, not speed. A model's weights take parameters times bytes per parameter, and the GPU needs a quarter to a half more on top for the KV cache and the runtime. The table applies that rule to each memory size on this page.
| Model size (weights) | RTX 3070, RTX 3070 Ti8 GB | RTX 308010 GB | RTX 3080 Ti12 GB | RTX 309024 GB |
|---|---|---|---|---|
| 7-8 billion parameters | 4-bit: Fits8-bit: No | 4-bit: Fits8-bit: Tight | 4-bit: Fits8-bit: Fits | 4-bit: Fits8-bit: Fits |
| 13-14 billion parameters | 4-bit: No8-bit: No | 4-bit: Tight8-bit: No | 4-bit: Fits8-bit: No | 4-bit: Fits8-bit: Fits |
| 30-34 billion parameters | 4-bit: No8-bit: No | 4-bit: No8-bit: No | 4-bit: No8-bit: No | 4-bit: Tight8-bit: No |
| 70 billion parameters | 4-bit: No8-bit: No | 4-bit: No8-bit: No | 4-bit: No8-bit: No | 4-bit: No8-bit: No |
It is the original RTX 3080 with 8,704 CUDA cores. NVIDIA later sold a 12 GB version with 8,960 cores under the same name, and many listings quote its memory. Size against 10 GB: an 8B model at 8-bit loads with little room left for context, and a 13B model at 8-bit does not load at all.
The RTX 3090 is the only RTX 30 GPU with 24 GB. Our larger Xeon machines take up to 3 of them, 72 GB in total, which llama.cpp or vLLM can split a model across. Two hold a 70B model at 4-bit, tightly. For one GPU with more memory, see every GPU we fit; for model-by-model plans, LLM servers.
Platforms
Every RTX 30 GPU has a PCIe 4.0 interface. On the Intel Xeon machines it runs at PCIe 3.0; the AMD EPYC machine gives the RTX 3080 a full PCIe 4.0 link. Once a model or a scene is loaded into GPU memory the link barely matters. It matters for training loops that stream batches from system memory every step.
| Machine | PCIe link | RTX 3070 | RTX 3070 Ti | RTX 3080 | RTX 3080 Ti | RTX 3090 |
|---|---|---|---|---|---|---|
| Intel Xeon E5-2600 v1/v2from $70.93/mo without a GPU | PCIe 3.0 | up to 1 | not offered | up to 2 | not offered | up to 2 |
| Intel Xeon E5-2600 v3/v4from $73.47/mo without a GPU | PCIe 3.0 | up to 2 | up to 2 | up to 2 | up to 2 | up to 3 |
| Intel Xeon Silver / Goldfrom $83.67/mo without a GPU | PCIe 3.0 | up to 2 | up to 2 | up to 2 | up to 2 | up to 3 |
| AMD EPYCfrom $217.59/mo without a GPU | PCIe 4.0 | not offered | not offered | up to 1 | not offered | not offered |
Machine prices are the cheapest build of each line without a GPU. RTX 30 servers can be ordered in New York, Miami, San Francisco and Bucharest. Amsterdam machines are not offered with these GPUs. Location sets latency to your users and the data-protection law that applies; data centers describes each building.
Workloads
One RTX 3090 serves a 13B model at 8-bit with room for context, or a 30B model at 4-bit with a short one, behind your own API with Ollama, llama.cpp or vLLM. Nothing you send leaves your machine.
Blender Cycles renders on CUDA or OptiX, and the RT cores accelerate ray tracing. Install the operating system your renderer needs and render overnight on a flat monthly price.
The GeForce driver runs DirectX, Vulkan and OpenGL as a desktop does. Run a game server, a cloud gaming host or a remote workstation with a dedicated GPU behind it.
Each GPU has one NVENC encoder for H.264 and HEVC and one NVDEC decoder. NVIDIA allows up to 8 encode sessions at once per system on GeForce GPUs, with driver 550.54.14 or later.
Delivery and terms
Nothing to fit: the server is installed and handed over. Instant servers lists what is racked right now.
A technician installs the GPU, builds the disk array and installs the operating system. Extra memory or other drives go in during the same visit.
Every timeline starts once payment and fraud review clear. Most orders clear within a couple of hours; a first order on a new account can take longer.
Owners also list RTX 30 machines on our marketplace, each at a monthly price the owner sets. Listings marked as coming through our sister company, Primcast, are rented on primcast.com.
Specifications
| GPU | Memory | Bus | Bandwidth | CUDA cores | Boost | Power |
|---|---|---|---|---|---|---|
| RTX 3070 | 8 GB GDDR6 | 256-bit | 448 GB/s | 5,888 | 1.73 GHz | 220 W |
| RTX 3070 Ti | 8 GB GDDR6X | 256-bit | 608 GB/s | 6,144 | 1.77 GHz | 290 W |
| RTX 3080 | 10 GB GDDR6X | 320-bit | 760 GB/s | 8,704 | 1.71 GHz | 320 W |
| RTX 3080 Ti | 12 GB GDDR6X | 384-bit | 912 GB/s | 10,240 | 1.67 GHz | 350 W |
| RTX 3090 | 24 GB GDDR6X | 384-bit | 936 GB/s | 10,496 | 1.70 GHz | 350 W |
Figures from NVIDIA's GeForce RTX 30 series comparison and its video encode and decode support matrix. Prices, machines, cities and per-machine counts are read from our order catalogue, the one the checkout bills from.
Answers
A complete RTX 3090 server starts at $309.93 a month. That is the GPU at $239 plus the cheapest machine that takes it: Intel Xeon E5-2630L, 6 cores, 2.00 GHz · 32 GB DDR3 · 2 × SATA 500 GB (RAID 1) · 500 Mbps. The cheapest RTX 30 build is the RTX 3070 at $209.93. A GPU fitted to order adds a one-time setup of $29.
The 10 GB version, with 8,704 CUDA cores and 760 GB/s of memory bandwidth. A complete RTX 3080 server starts at $229.93 a month. If your model needs 12 GB, take the RTX 3080 Ti, which has 12 GB on a 384-bit bus.
Yes. Our larger Xeon machines take up to 3 RTX 3090s, 72 GB of GPU memory in total, and the configurator shows the most each machine takes. Each GPU is its own monthly line, so the price grows by the GPU's figure for each one you add.
For inference and QLoRA fine-tuning, yes. Its 24 GB holds a 13B model at 8-bit with room for context, or a 30B model at 4-bit with a short one. It has no ECC memory, so for long training runs without checkpoints a compute GPU is the safer choice.
About 30 minutes after your order is approved when a machine with that GPU is already racked; those are listed on the instant servers page. When we fit the GPU to the machine you configured, allow 4 to 24 hours.
No. The machine is rented to one account and the GPU is passed straight through to your operating system. There is no hypervisor, no vGPU profile and no time-slicing, and you install the NVIDIA driver and CUDA version you want.
No. RTX 30 GPUs decode AV1 but encode only H.264 and HEVC, on one NVENC encoder per GPU. Hardware AV1 encoding starts with the RTX 40 series.
In New York, Miami, San Francisco and Bucharest, on Server Room's own network. The machines that take these GPUs are not offered with them in Amsterdam.
Choose the GPU by memory, then the processor, memory, disks and city. The configurator prices every change as you make it.
Every GPU we fit · The RTX 3050 desktop · RTX 40 servers · Racked right now