NVIDIA RTX 5090 Hosting on Bare Metal in the UK
Single-tenant servers built around NVIDIA’s flagship Blackwell consumer GPU. 32 GB of GDDR7 memory, 1,792 GB/s memory bandwidth, 21,760 CUDA cores, hardware FP4 acceleration. The fastest single GPU we’ll rent you.
RTX 5090 Server Specs
The hardware you actually rent.
| GPU model | NVIDIA GeForce RTX 5090 (Blackwell, GB202) |
|---|---|
| Architecture | Blackwell — 5th gen Tensor Cores, 4th gen RT Cores |
| VRAM | 32 GB GDDR7 @ 1,792 GB/s |
| CUDA cores | 21,760 |
| FP16 compute | ~ 105 TFLOPS |
| FP8 / FP4 throughput | ~ 838 / ~ 1,676 TOPS (sparsity off) |
| TDP | 575 W |
| Host CPU | AMD Ryzen 9 / Threadripper (16+ cores) |
| Host RAM | Up to 128 GB DDR5 |
| Storage | 1.92 TB NVMe + optional 8 TB SATA SSD |
| Network | 1 Gbps unmetered, optional 10 Gbps |
| Location | London, United Kingdom |
What Fits on a Single RTX 5090
32 GB is a sweet spot — comfortable for 13B FP16, generous for 7B with long context, and capable of 32B/70B at INT4. The chart below covers what we benchmark in production.
| Model | Params | FP16 | INT4 / FP8 | Notes |
|---|---|---|---|---|
| Mistral 7B Instruct | 7B | 14 GB FP16 | 5 GB INT4 | Easy fit at FP16 with 32K context |
| LLaMA 3.1 8B | 8B | 16 GB FP16 | 5 GB INT4 | Fits with full 128K context at FP16 |
| Qwen 2.5 14B | 14B | 28 GB FP16 | 9 GB INT4 | Fits FP16 — typical chatbot tier |
| Llama 3.3 70B | 70B | 140 GB FP16 | 40 GB INT4 | INT4/AWQ only on a single 5090 |
| DeepSeek-V2 16B | 16B | 32 GB FP16 | 10 GB INT4 | Tight at FP16, comfy at INT4 |
| Gemma 2 27B | 27B | 54 GB FP16 | 16 GB INT4 | INT4 fits, FP16 needs 6000 Pro |
| FLUX.1 dev | 12B | 24 GB | 12 GB FP8 | Native FP8 path on Blackwell |
| SDXL 1.0 | 3.5B | 8 GB FP16 | 4 GB FP8 | Trivial — runs at full speed |
| Whisper Large-v3 | 1.5B | 6 GB FP16 | n/a | Real-time transcription, batch 8+ |
| HunyuanVideo | 13B | 30 GB FP8 | n/a | FP8 fits, FP16 needs 6000 Pro |
Why People Pick the RTX 5090
Real customer workloads we run on this hardware every day.
Production chatbot APIs
vLLM on a 5090 hits ~1,200 tok/s aggregate on Mistral 7B FP16 — enough for hundreds of concurrent users on a single card. The Blackwell FP8 path drops to ~£0.04 per million tokens at full utilisation.
FLUX.1 and SDXL hosting
Native FP8 on Blackwell halves FLUX.1 memory pressure vs Ada. Generate 1024×1024 in ~6 seconds, batch 4 simultaneously without OOM.
AI video generation
13B-class video models (Wan 2.1, HunyuanVideo, LTX-Video) fit at FP8. Render a 5-second clip in ~3-5 minutes — not real-time, but practical for production teams.
Real-time speech
Whisper Large-v3 + a TTS model on the same card. Process 16 concurrent voice agents with sub-300ms time-to-first-audio.
Fine-tuning
QLoRA on 13B–32B models is comfortable. Full SFT of 7B models trains overnight. ZeRO-2 hands you 70B fine-tuning if you pair two 5090s.
Research workloads
21,760 CUDA cores and 32 GB VRAM is enough for most academic NLP and CV experiments. Cheaper than an A100 40 GB and faster on FP4/FP8.
RTX 5090 vs the Rest of the Catalogue
How this card stacks up against the rest of the GigaGPU catalogue for the workloads we benchmark.
| GPU | VRAM | Throughput / Notes | 70B INT4 fits? | Price | |
|---|---|---|---|---|---|
| RTX 5090 | 32 GB GDDR7 | ~1,200 | ~6 s | Yes (tight) | from £359 |
| RTX 6000 Pro 96 GB | 96 GB GDDR7 | ~1,150 | ~6 s | Yes (FP16) | from £1,099 |
| RTX 5080 | 16 GB GDDR7 | ~1,050 | ~7 s | No | from £189 |
| RTX 4090 | 24 GB GDDR6X | ~950 | ~8 s | Yes (tight) | from £279 |
| RTX 3090 | 24 GB GDDR6X | ~720 | ~14 s | Yes (tight) | from £159 |
| A100 80 GB | 80 GB HBM2e | ~1,300 | ~9 s | Yes (FP16) | POA |
Deep Dive
Why 32 GB matters more than headline tok/s
Benchmarks on Mistral 7B make the 5090 look 1.6× faster than a 3090. That’s true. What’s more important: 32 GB lets you run 13B FP16 with a 32K-token context window, and 70B INT4 with vLLM’s PagedAttention — neither of which fits on a 24 GB card without aggressive offloading. Memory unlocks which models you can serve, not just how fast.
Blackwell FP4 — when it pays off
The Blackwell architecture introduces hardware FP4 (4-bit floating point) for the first time on a consumer card. For inference workloads using NVFP4 or MX-FP4 quantisation (Llama 3 8B, Mistral 7B), throughput is roughly 2× FP8 with a quality drop of <1% on standard benchmarks. Frameworks worth knowing about: TensorRT-LLM, vLLM 0.6+, NVIDIA NIM, and SGLang.
FP4 is not magic — it doesn’t help training, it doesn’t help diffusion models that haven’t been quantised, and most fine-tunes will still ship in BF16. But for serving production LLMs at >1K tok/s on a single GPU, it’s the cheapest performance jump in years.
How the 5090 compares to a 6000 Pro for serving
The RTX 6000 Pro 96 GB Workstation is the same Blackwell silicon as the 5090 with three differences: 3× the VRAM (96 GB vs 32 GB), ECC memory, and certified pro drivers. For raw token throughput on models that fit on the 5090, throughput is within 5%. The Pro card earns its keep when:
- You need to serve 70B FP16 on a single card (not INT4)
- You’re running long-context (>32K) on a 32B+ model
- ECC matters for compliance or research reproducibility
- You’re stacking multiple models on one GPU and the 32 GB is constraining
For cost-per-token on common chatbot workloads (7B–14B FP16/FP8), the 5090 wins. For headroom on bigger models, take the 6000 Pro.
Frequently Asked Questions
The questions buyers actually ask before committing to a GPU server.
Is this the consumer GeForce RTX 5090 or the workstation card?
It’s the consumer GeForce RTX 5090, founders edition or AIB depending on stock. The professional Blackwell variant is the RTX 6000 Pro 96 GB.
Can I run two 5090s in one server?
Yes — see our multi-GPU clusters. 2× RTX 5090 in tensor-parallel handles Llama 3 33B FP16 or 70B INT4 with headroom. There’s no NVLink — they communicate over PCIe Gen 4 x16.
How does it compare to the RTX 4090?
Blackwell vs Ada: ~25% faster on FP16, ~2× faster on FP8/FP4, and 32 GB vs 24 GB VRAM. The 5090 is the better choice for almost any new deployment unless 4090 stock is genuinely cheaper. See our comparison breakdown.
What’s the power draw?
575 W TDP at peak. We provision N+1 power per rack at 30°C cold-aisle and have validated continuous 100% load on every node — no thermal throttling on workloads we benchmarked.
Is FP4 supported in vLLM?
Yes from v0.6+ with the right model checkpoint (NVFP4 or MX-FP4 quantised). Mistral 7B and Llama 3.1 8B have FP4 community ports on Hugging Face.
How quickly can I get one?
In-stock SKUs ship same-day to next-day. Multi-GPU 5090 clusters typically ship in 3–5 working days.
Does it run Stable Diffusion 3 / FLUX.1?
Yes — see our Can the RTX 5090 run FLUX? page for the full benchmark.
Can I run Whisper and an LLM on the same card?
Yes. 32 GB comfortably fits Whisper Large-v3 (6 GB) plus a 7B LLM at FP16 (14 GB). Concurrent serving works well with separate vLLM and faster-whisper processes.
Related Pages
Pages our visitors typically read next.
Deploy a dedicated RTX 5090 today.
32 GB of GDDR7, FP4 acceleration, single-tenant bare metal in the UK. From £399/mo with same-day deployment on in-stock cards.
