← All categories
Hybrid GPU + RAM
A GPU for attention and context with large system RAM holding expert weights: the usual way to run very large MoE models without many GPUs.
| Machine | Memory for models | Bandwidth | Price | Largest at 4-bit | gpt-oss-20b |
|---|---|---|---|---|---|
| RTX 5090 + EPYC 9355P with 768 GB DDR5 (MoE offload)Hybrid GPU + RAM | 32 GB | 1,792 GB/s | ~$3,763 + unpriced partsGPU, motherboard, ram and cooling prices not included | Nemotron 3.5 Lightning 30B-A3B | 290 tok/s |
| RTX 3090 (used) + 128 GB DDR5 on AM5 (MoE offload)Hybrid GPU + RAM | 24 GB | 936 GB/s | ~$4,833 + unpriced partsused GPU price not included | Qwen3.6 35B-A3B | 162 tok/s |