Table of Contents
Companion to RTX 4090 spec breakdown for the Blackwell flagship.
TL;DR
RTX 5090 = 32 GB GDDR7, 21,760 CUDA cores, 1,792 GB/s bandwidth, native FP8 (~838 TOPS) and FP4 (~1,676 TOPS). The fastest single GPU we host. £399/mo. Best cost-per-token at FP8 in our catalogue.
Full spec sheet
| Spec | RTX 5090 |
|---|---|
| Architecture | Blackwell GB202 |
| VRAM | 32 GB GDDR7 |
| Memory bus | 512-bit |
| Memory bandwidth | 1,792 GB/s |
| CUDA cores | 21,760 |
| Tensor cores (5th gen) | 680 |
| FP16 TFLOPS | ~210 |
| FP8 TOPS | ~838 |
| FP4 TOPS | ~1,676 |
| TDP | 575 W |
| PCIe | Gen 5 x16 |
| Launch year | 2025 |
| Monthly (GigaGPU) | £399 |
AI relevance
- 32 GB enables Llama 3 8B FP16 + 32K context, Qwen 2.5 14B FP16, 70B INT3
- FP8 hardware = 50% throughput uplift over FP16
- FP4 hardware (NVFP4 / MX-FP4) = additional 2× over FP8 on supported models
- 1,792 GB/s bandwidth = best-in-class for memory-bound LLM inference
Comparisons
See vs RTX 3090, vs RTX 4090, vs 6000 Pro.
Verdict
The RTX 5090 is the best price-per-performance AI GPU we rent. For new 2026 deployments it's the default flagship.
Bottom line
RTX 5090 = best per-pound flagship. See 5090 hosting page.
