RTX 5090 vs RTX 4090 für lokale KI: welche 2026 wählen?
Der nüchterne Vergleich aus dem Blickwinkel, der zählt: welche Modelle laufen, bei welchem Durchsatz, zu welchem Preis. VRAM (32 vs 24 GB), GDDR7-Bandbreite, Modelltabelle und Urteil nach Profil.
Uebersetzter Artikel. Diese Version ist lokalisiert, damit internationale Seiten keinen franzoesischen Artikeltext anzeigen. Technische Daten, Preise und Empfehlungen bleiben gleich.
Key numbers
| VRAM | 24 GB GDDR6X | 32 GB GDDR7 |
| Memory bandwidth | ~1,008 GB/s | ~1,792 GB/s |
| TDP | 450 W | 575 W |
| Indicative new price 2026 | ~EUR 1,600-1,900 | ~EUR 2,200-2,700 |
The number that changes everything
The RTX 5090 has about 78% more memory bandwidth thanks to GDDR7. LLM inference is memory-bound: tokens per second depend far more on bandwidth than on raw compute. That, plus the 8 extra gigabytes, is where the 5090 pulls ahead.
Which models run on each
| Mistral 7B / Llama 8B | FP16 yes | FP16 yes |
| Qwen 32B / Gemma 31B | Q4 yes, Q5 tight | Q5 yes, Q6 yes |
| Llama 3.3 70B / Qwen 72B | Q3 very tight | Q3 yes (Q4 out of reach) |
| Llama 4 Scout 109B (MoE) | Too big | Needs two cards |
Throughput and value
On a model that fits both cards, the RTX 5090 delivers roughly 40 to 70% more tokens per second. The gap is most visible on multi-request use and long contexts.
The used RTX 4090 (around EUR 1,100-1,400) stays the best value if your models fit in 24 GB, you run solo or in a small team, and you do not need the 70B tier. The 5090 is worth it for 70B on a single card, long context, or serving several users.
- Solo dev, models up to 32B: used RTX 4090 — unbeatable value.
- You want Llama 70B on one card: RTX 5090 — the 4090 cannot follow.
- Multi-user or long context: RTX 5090 — the GDDR7 bandwidth pays off.
- 100B+ models: multi-GPU (NVLink) — neither card alone.
Enter your card and target model in the LocalIA GPU to LLM calculator: it tells you whether it fits, at which quantization, with estimated VRAM and throughput, across 200+ GPUs. Free, no signup. LocalIA is an independent resource and sells nothing.
Rechner öffnen / frag uns um Rat mit Zielmodell, Nutzern und Randbedingungen.