Table of Contents
The 4090 launched in 2022 and ruled the consumer flagship slot until the 5090 in early 2025. For new AI deployments in 2026, the 5090 wins on most workloads — especially anything FP8.
RTX 5090 is ~30% faster on FP16 and ~2× faster on FP8 vs RTX 4090. 32 GB vs 24 GB makes 14B FP16 fit. Pricing 1.3× (£289 vs £289/mo). For new deployments the 5090 is the right pick; the 4090 only wins when stock pricing is dramatically cheaper.
Spec deltas
| Spec | RTX 4090 | RTX 5090 | Delta |
|---|---|---|---|
| Architecture | Ada Lovelace | Blackwell | +1 gen |
| VRAM | 24 GB GDDR6X | 32 GB GDDR7 | +33% |
| Memory bandwidth | 1,008 GB/s | 1,792 GB/s | +78% |
| CUDA cores | 16,384 | 21,760 | +33% |
| FP16 TFLOPS | ~165 | ~210 | +27% |
| FP8 hardware | No | Yes (~838 TOPS) | ∞ |
| FP4 hardware | No | Yes (~1,676 TOPS) | ∞ |
| TDP | 450 W | 575 W | +28% |
| Monthly (GigaGPU) | £289 | £399 | +29% |
Real benchmarks
| Workload | RTX 4090 | RTX 5090 | Speedup |
|---|---|---|---|
| Mistral 7B FP16 (aggregate) | 950 tok/s | 1,180 tok/s | 1.24× |
| Mistral 7B FP8 (aggregate) | sw fp8 ~960 | 1,920 | 2.0× |
| Llama 3 8B FP8 | sw ~920 | 1,820 | 2.0× |
| Qwen 2.5 14B FP16 | tight, OOM possible | 720 (fits) | ∞ |
| SDXL 1024² (s/img) | 4 s | 2.5 s | 1.6× |
| FLUX.1 dev FP8 (s/img) | 8 s sw | 6 s native | 1.33× |
| Whisper Large-v3 RTF | ~7× | ~9× | 1.29× |
Pricing
5090 is £80/mo more than 4090. Per-token cost-per-1M-tokens:
- Mistral 7B FP16: 4090 £0.19, 5090 £0.20 (essentially tied)
- Mistral 7B FP8: 4090 ~£0.19, 5090 £0.12 (5090 wins decisively)
Verdict
For new 2026 deployments the 5090 is the right pick — better cost-per-token at FP8, 32 GB unlocks 14B FP16, futureproof on quantisation. The 4090 only wins on Ada-only workloads or when stock pricing is significantly cheaper.
Bottom line
5090 wins on capability and cost-per-token at FP8. See RTX 5090 vs RTX 3090 for the older comparison.
