2x NVIDIA RTX 5090
DesktopTwo RTX 5090s, 64 GB of VRAM in total. A multi-GPU rig is not a single memory pool, so bandwidth and TDP are left empty.
7
Run records
6
Models covered
2
Frameworks covered
L0 Self-reported × 7
Evidence mix
Typical measured performance
Published benchmark averages per model × framework × quantization on this hardware.
- vLLM · IQ3_S Decode — · Prefill — ·
- vLLM · IQ2_XS Decode — · Prefill — ·
- vLLM · IQ3_S Decode — · Prefill — ·
- vLLM · IQ3_S Decode — · Prefill — ·
- llama.cpp · UD-Q8_K_XL Decode 32.08 · Prefill 671.45 ·
- llama.cpp · UD-IQ4_XS Decode 21.96 · Prefill 421.07 ·
- llama.cpp · UD-IQ2_XXS Decode 19.39 · Prefill 692.41 ·
Fit check
Physics-based estimate from VRAM and memory bandwidth: pick a model to see which quantizations fit and the expected speed
Formula: tok/s = (bandwidth ÷ model size) × 0.55, adapted from llmfit; see run records for measured numbers
Key specs
- Vendor
- NVIDIA
- Architecture
- Blackwell
- VRAM
- 64 GB
- Memory bandwidth
- 1792 GB/s
Run records
7 configurations
| Model | Framework | Quantization | Decode | Gen time | Samples |
|---|---|---|---|---|---|
| Laguna-S-2.1 | vLLM | IQ3_S | — | — | 1 |
| Laguna-S-2.1 | vLLM | IQ2_XS | — | — | 1 |
| GLM-5.3 | vLLM | IQ3_S | — | — | 1 |
| DeepSeek-V4.1-Flash | vLLM | IQ3_S | — | — | 1 |
| DeepSeek-V4-Flash-Vision-Exp | llama.cpp | UD-Q8_K_XL | 32.1 tok/s | — | 1 |
| GLM-5.3-Flash | llama.cpp | UD-IQ4_XS | 22 tok/s | — | 1 |
| DeepSeek-V4-Flash | llama.cpp | UD-IQ2_XXS | 19.4 tok/s | — | 1 |