2x RTX 3090 (48GB total) vs RTX 5090 32GB for local LLMs
Memory fit, speed and context compared with the same math as the advisor. 2x RTX 3090 (48GB total) has 15.0 GB more usable memory (46.0 vs 31.0 GB) and higher bandwidth than the other (936 vs 1792 GB/s).
Verdict from the numbers
- Fit. 2x RTX 3090 (48GB total) fits 11 of 13 models fully at 16k context (largest: Qwen3.6 35B-A3B Q8_0); RTX 5090 32GB fits 11 (largest: Qwen3.6 35B-A3B UD-Q6_K). Both fit the same set of models.
- Speed. On the 11 models that fit fully on both, RTX 5090 32GB is faster by a typical 1.5x. 2x RTX 3090 (48GB total) is ahead on 2, RTX 5090 32GB on 9. No model is measured on both machines, so every comparison involves an estimate; read it as a guide.
- Trade-off. 2x RTX 3090 (48GB total) has 15.0 GB more usable memory (46.0 vs 31.0 GB) and higher bandwidth than the other (936 vs 1792 GB/s).
Where to check prices
Price tier note
We do not list prices because they move daily and differ between new and used stock. What the data does show is what the memory difference buys: 2x RTX 3090 (48GB total) has 15.0 GB more usable memory, which moves 0 of 13 models into full-GPU fit. Use the Amazon links above to compare current listings and decide whether the difference is worth it for the models you want to run.
Model by model · 16k context, f16 KV cache
| Model | 2x RTX 3090 (48GB total) fit | tok/s | RTX 5090 32GB fit | tok/s | Faster / fits |
|---|---|---|---|---|---|
| Gemma 4 26B-A4B Q8_0 | fits fully | 125 estimated | fits fully | 213 scaled from measured | RTX 5090 32GB |
| Gemma 4 31B Q8_0 | fits fully | 19 estimated | fits fully | 55 measured | RTX 5090 32GB |
| Gemma 4 12B Q8_0 | fits fully | 49 estimated | fits fully | 91 scaled from measured | RTX 5090 32GB |
| Qwen3.6 27B Q8_0 | fits fully | 22 estimated | fits fully | 49 scaled from measured | RTX 5090 32GB |
| Qwen3.6 35B-A3B Q8_0 | fits fully | 163 estimated | fits fully | 260 measured | RTX 5090 32GB |
| Qwen3.5 9B Q8_0 | fits fully | 67 estimated | fits fully | 110 scaled from measured | RTX 5090 32GB |
| Qwen3.5 4B Q8_0 | fits fully | 143 estimated | fits fully | 72 scaled from measured | 2x RTX 3090 (48GB total) |
| gpt-oss 20B Q8_0 | fits fully | 241 estimated | fits fully | 298 scaled from measured | RTX 5090 32GB |
| gpt-oss 120B Q4_K_M | MoE offload to CPU | 55 estimated | does not fit | - | - |
| Qwen3-Coder 30B-A3B Q8_0 | fits fully | 143 estimated | fits fully | 316 scaled from measured | RTX 5090 32GB |
| Mistral Small 3.2 24B Q8_0 | fits fully | 26 estimated | fits fully | 56 scaled from measured | RTX 5090 32GB |
| Qwen3-VL 8B Q8_0 | fits fully | 74 estimated | fits fully | 43 scaled from measured | 2x RTX 3090 (48GB total) |
| GLM-4.5-Air Q4_K_M | MoE offload to CPU | 14 estimated | does not fit | - | - |
Badges: measured medians of real llama.cpp runs; scaled from measured and estimated come from our bandwidth model. See sources and methods.
How much context fits?
Largest context where weights plus KV cache stay inside usable memory, per model that fits fully on at least one side. q8_0 KV roughly halves the cache.
| Model | 2x RTX 3090 (48GB total) f16 | RTX 5090 32GB f16 | 2x RTX 3090 (48GB total) q8_0 | RTX 5090 32GB q8_0 |
|---|---|---|---|---|
| Qwen3.6 35B-A3B | 256k | 64k | 512k | 128k |
| Gemma 4 31B | 128k | 32k | 256k | 64k |
| Qwen3-Coder 30B-A3B | 128k | 32k | 128k | 64k |
| Qwen3.6 27B | 256k | 32k | 256k | 32k |
| Gemma 4 26B-A4B | 256k | 128k | 256k | 256k |
| Mistral Small 3.2 24B | 64k | 32k | 128k | 64k |
| gpt-oss 20B | 128k | 128k | 128k | 128k |
| Gemma 4 12B | 256k | 256k | 256k | 256k |
| Qwen3.5 9B | 1M | 512k | 1M | 1M |
| Qwen3-VL 8B | 128k | 128k | 256k | 256k |
| Qwen3.5 4B | 1M | 512k | 1M | 1M |
Spec summary
| 2x RTX 3090 (48GB total) | RTX 5090 32GB | |
|---|---|---|
| Usable memory for model + KV cache | 46.0 GB | 31.0 GB |
| Memory bandwidth | 936 GB/s | 1792 GB/s |
| Models that fit fully at 16k | 11 of 13 | 11 of 13 |
| Models that need offload or do not fit | 2 | 2 |
| Measured data points on file | none (multi-GPU is estimated) | 38 |
FAQ
Which is faster for local LLMs, 2x RTX 3090 (48GB total) or RTX 5090 32GB?
On the 11 models that fit fully on both, RTX 5090 32GB is typically 1.5x faster (geometric mean). 2x RTX 3090 (48GB total) wins 2 and RTX 5090 32GB wins 9. 0 comparisons are measured on both.
Which fits more models, 2x RTX 3090 (48GB total) or RTX 5090 32GB?
2x RTX 3090 (48GB total) fits 11 of 13 models fully at 16k context and RTX 5090 32GB fits 11.
Can I run long context on 2x RTX 3090 (48GB total) or RTX 5090 32GB?
At f16 KV cache, the largest-model context limits are in the table above (for example Qwen3.6 35B-A3B on 2x RTX 3090 (48GB total): 256k; Qwen3.6 35B-A3B on RTX 5090 32GB: 64k). A q8_0 KV cache extends it further.
Do you state prices for 2x RTX 3090 (48GB total) or RTX 5090 32GB?
No. Prices change daily, so we link to Amazon search results instead and focus on measured and estimated speed and memory fit.
Related
2x RTX 3090 (48GB total) · RTX 5090 32GB · Measured model pages: Qwen3.6 27B
More comparisons: 2x RTX 3090 (48GB total) vs RTX PRO 6000 96GB (bonus page) · Mac Studio (2025) M3 Ultra 96GB vs 2x RTX 3090 (48GB total) · RTX 4090 24GB vs RTX 5090 32GB · RTX 3090 24GB vs RTX 5090 32GB · all comparisons
By memory: 8 GB · 12 GB · 16 GB · 24 GB · 32 GB · 48 GB · 64 GB · 96 GB · 128 GB · 192 GB · 256 GB · 512 GB