Table of Contents
Per-pound efficiency is the right metric for cost-sensitive deployments. RTX 3090 at £159/mo and RTX 5090 at £399/mo trade off across different workloads. This page is the numerical breakdown.
RTX 3090 wins per-pound on FP16 inference of small models. RTX 5090 wins decisively on any FP8 workload (most modern LLMs ship FP8 checkpoints) and on workloads that need 24+ GB. For a new deployment in 2026 the 5090 is the better choice 80% of the time.
The price/throughput ratio
| Workload | RTX 3090 tok/s | RTX 5090 tok/s | Speedup | Cost ratio (5090/3090) | Per-pound winner |
|---|---|---|---|---|---|
| Mistral 7B FP16 | 720 | 1,180 | 1.64× | 2.0× | 3090 |
| Mistral 7B FP8 | n/a | 1,920 | ∞ | n/a | 5090 |
| Llama 3 8B FP16 | 680 | 1,140 | 1.68× | 2.0× | 3090 (slightly) |
| Llama 3 8B FP8 | n/a | 1,820 | ∞ | n/a | 5090 |
| Qwen 2.5 14B FP16 | OOM | 720 | ∞ | n/a | 5090 |
| Mixtral 8x7B INT4 | OOM | ~280 | ∞ | n/a | 5090 |
| SDXL (s/image) | 14 s | 6 s | 2.33× | 2.0× | 5090 |
| FLUX.1 dev FP8 | sw fp8 ~16s | native ~6s | 2.67× | 2.0× | 5090 |
| Whisper Large-v3 RTF | 6× | 9× | 1.5× | 2.0× | 3090 |
By workload
- FP16 7B chatbot: 3090 wins by ~20% on cost-per-token.
- FP8 7B chatbot: 5090 wins decisively (3090 has no FP8 hardware).
- 14B+ models: 5090 wins (3090 cannot fit at FP16 with KV cache).
- Image generation: 5090 wins (newer arch + FP8 path).
- Whisper / embeddings only: 3090 wins (cheapest 24 GB card).
Why FP8 changes everything
The 5090’s hardware FP8 path delivers ~50% more throughput than FP16 at <1% quality cost. The 3090 has no hardware FP8 — software FP8 emulation is roughly the same speed as FP16. So:
- 3090 FP16 vs 5090 FP16: 5090 is 1.6× faster at 2× the cost — 3090 wins per-pound.
- 3090 FP16 vs 5090 FP8: 5090 is 2.7× faster at 2× the cost — 5090 wins per-pound.
Almost every modern LLM (Llama 3.x, Mistral, Qwen 2.5, FLUX.1) ships an FP8 checkpoint. Locking yourself out of FP8 is locking yourself into worse cost-per-token.
Verdict
For new 2026 deployments, the RTX 5090 wins on per-pound throughput on most workloads thanks to FP8. The RTX 3090 remains the right pick when 1) FP16 is your only target, 2) the model fits 24 GB comfortably, 3) cost is the dominant factor.
Bottom line
The 3090 is the cheapest 24 GB card we host and remains a good pick for FP16-only workloads. The 5090 is the better value for new deployments. See RTX 5090 vs RTX 3090 for the broader spec comparison.
