NVIDIA GeForce RTX 5090
Eight GeForce RTX 5090 GPUs with 32 GB of GDDR7 each, 256 GB per server, for diffusion image and video pipelines, rendering and inference of smaller models, configured on request.
- GPUs per server
- 8
- Memory per GPU
- 32 GB GDDR7
- Memory bandwidth
- 1,792 GB/s
- FP4 AI TOPS
- 3,352 (sparse)
Specifications
| GPU | |
|---|---|
| Architecture | NVIDIA Blackwell (GB202) |
| Memory | 32 GB GDDR7, 512-bit, 28 Gbps |
| Memory bandwidth | 1,792 GB/s |
| CUDA cores | 21,760 |
| Tensor Cores | 680, fifth generation |
| RT Cores | 170, fourth generation, 318 TFLOPS |
| FP4 Tensor (AI TOPS) | 3,352 TFLOPS sparse, 1,676 TFLOPS dense |
| FP8 Tensor, FP16 accumulate | 1,676 TFLOPS sparse, 838 TFLOPS dense |
| FP32 | 104.8 TFLOPS |
| Boost clock | 2.41 GHz |
| Video engines | 3 NVENC (ninth generation), 2 NVDEC (sixth generation), AV1 encode and decode |
| Interconnect | PCIe Gen 5; no NVLink |
| Total graphics power | 575 W |
| Reference card | 2-slot, 304 mm x 137 mm (Founders Edition) |
| Server as delivered | |
|---|---|
| GPUs per server | 8 |
| GPU memory per server | 256 GB (8 x 32 GB) |
| GPU power per server | 4,600 W at NVIDIA's rating (8 x 575 W) |
| Card model | Confirmed in your proposal |
| CPU, system memory, storage and network | Configured to your requirements and confirmed in your proposal |
Typical workloads
- Diffusion image and video generation, such as ComfyUI pipelines
- Inference of models that fit in 32 GB per GPU
- GPU rendering and real-time graphics
- Video encoding with AV1
Rent dedicated servers with eight NVIDIA GeForce RTX 5090 GPUs for generative image and video, rendering and inference of smaller models. RTX 5090 servers are available on request, configured to your requirements, on 36, 48 or 60 month contracts.
Who the RTX 5090 is for
The RTX 5090 is for pipelines that run one job per GPU and need fast memory more than large memory: diffusion image and video generation, GPU rendering, and inference of models that fit in 32 GB. Each GPU has 32 GB of GDDR7, 256 GB per server, on the Blackwell architecture with FP4 support in its fifth-generation Tensor Cores.
Teams that render or generate frames in parallel get eight independent GPUs per server, each with the full memory and bandwidth of the card and no other tenant on the machine.
What it does well
Memory bandwidth is the standout. NVIDIA rates the RTX 5090 at 1,792 GB/s on a 512-bit bus at 28 Gbps, more per GPU than the 1,597 GB/s of the RTX PRO 6000 Blackwell Server Edition.
- Compute: 21,760 CUDA cores and 104.8 TFLOPS of FP32. NVIDIA's 3,352 AI TOPS is FP4 with sparsity; dense FP4 is 1,676 TFLOPS.
- Tensor and RT Cores: 680 fifth-generation Tensor Cores and 170 fourth-generation RT Cores, rated at 318 TFLOPS for ray tracing.
- Video: three ninth-generation NVENC and two sixth-generation NVDEC engines per GPU, with AV1 encode and decode.
- Power: 575 W of total graphics power per card at NVIDIA's rating, 4,600 W across eight GPUs.
The RTX 5090 has no NVLink. GPUs communicate over PCIe Gen 5, which suits work that runs independently on each GPU, such as one diffusion job per card, more than models split across GPUs.
Configuration as delivered
RTX 5090 servers are configured to your requirements and confirmed in your proposal, including the card model, CPU, system memory, storage and network.
Every server is burned in before handover, and the results go into an acceptance report, as our burn-in and acceptance page describes.
How renting works
RTX 5090 servers are available on request. Tell us how many servers you need, for how long and from when, and we confirm the delivery schedule in a written proposal. You rent whole 8-GPU servers on a 36, 48 or 60 month contract, and longer terms carry lower rates.
The company that will run the servers completes three onboarding forms before signing. A down payment at signing is followed by one fixed monthly rate for the full term, with no egress fees on the media you move. We do not offer trials; burn-in and the acceptance report take their place. Our how to buy page explains each step.
Workloads and alternatives
The RTX 5090 fits ComfyUI and similar diffusion pipelines, rendering, and inference where weights and KV cache stay within 32 GB per GPU. A 24B-parameter model in FP8 takes about 24 GB, which leaves about 8 GB on the card for the KV cache and working memory.
Our generative video page covers diffusion pipelines on dedicated servers. The real cost of running RTX 5090 at scale looks at data movement and idle time, and ComfyUI and AI video on dedicated bare metal covers the pipeline itself.
When 32 GB is too small for your resolution, clip length or batch size, the RTX PRO 6000 Blackwell has 96 GB per GPU with ECC. For pipelines tuned for Ada Lovelace, see the RTX 4090.
Questions buyers ask
RTX 5090 or RTX PRO 6000 Blackwell: which should we rent?
The RTX PRO 6000 Blackwell has three times the memory per GPU (96 GB against 32 GB), ECC and a passively cooled server design, and allocation is open in Montreal. The RTX 5090 has more memory bandwidth per GPU, 1,792 GB/s against 1,597 GB/s, and suits work that fits in 32 GB.
What does 3,352 AI TOPS mean?
It is NVIDIA's peak FP4 Tensor figure for the RTX 5090 with sparsity, from its RTX Blackwell architecture whitepaper. Without sparsity, the figure is 1,676 TFLOPS.
Are the eight GPUs linked with NVLink?
No. The GeForce RTX 5090 has no NVLink, so the GPUs communicate over PCIe Gen 5. Pipelines that run an independent job on each GPU suit it best.
Which RTX 5090 cards are in the servers?
The card model is confirmed in your proposal. NVIDIA's figures on this page describe its Founders Edition or reference design, and NVIDIA notes that board partner cards can differ.
What contract terms are available?
36, 48 or 60 months, with lower rates for longer terms. You pay a down payment at signing, then one fixed monthly rate for the full term, with no egress fees.
Rent RTX 5090 servers on 36 to 60 month terms.
Share the hardware, the number of servers, the term and the start date. You receive a written proposal built on our delivery specification.