The mainstream rack-server inference card of the Ada generation: 48 GB GDDR6, passive cooling (needs server chassis airflow), widely available in 1–8× configurations from Dell, HPE, Lenovo, and Supermicro. A common building block for on-prem AI pilots because it slots into standard 2U servers without exotic power or cooling.
| VRAM | 48 GB |
| Memory Bandwidth | 864 GB/s |
| TDP | 350 W |
| Architecture | Ada Lovelace AD102 |
| Release Year | 2023 |
| MSRP at Launch | $7,499 |
| Inference Speed (Llama 3.1 8B Q4_K_M) | ~130 tokens/sec |
| Inference Speed (Llama 3.3 70B Q4_K_M) | ~24 tokens/sec |
All models below run comfortably in 48 GB VRAM with Q4_K_M quantization.
| Llama 4 | 67 GB VRAM (smallest variant — needs more VRAM or a lower quant) · Q4_K_M · ollama run llama4:scout |
| Llama 3.3 | 43 GB VRAM · Q2_K_XS (Tight) · ollama run llama3.3 |
| Llama 3.1 Family | 6 GB VRAM · Q4_K_M · ollama run llama3.1 |
| DeepSeek R1 | 20 GB VRAM · Q4_K_M · ollama run deepseek-r1:32b |
| Qwen 3 | 20 GB VRAM · Q4_K_M · ollama run qwen3:32b |
| Qwen 3.5 | 22 GB VRAM · Q4_K_M · ollama run qwen3.5:35b-a3b |
| Qwen 3.6 | 22 GB VRAM · Q4_K_M · ollama run qwen3.6:35b-a3b |
| Gemma 3 | 17 GB VRAM · Q4_K_M · ollama run gemma3:27b |
Install Ollama then run the recommended model for this GPU:
ollama run llama3.3:70b
Yes — the NVIDIA L40S has 48 GB VRAM and runs The mainstream rack-server inference card of the Ada generation: 48 GB GDDR6, passive cooling (needs server chassis airf
The NVIDIA L40S runs Llama 3.1 8B at ~130 tokens/sec with Q4_K_M quantization. For the 70B model it achieves ~24 tokens/sec.
With 48 GB you can run: Llama 4, Llama 3.3, Llama 3.1 Family, DeepSeek R1, Qwen 3. Use Ollama for the easiest setup: ollama run llama3.3:70b.
← All GPU Reviews | Check Your Hardware | Full Benchmarks | Can I Run It?