Can RTX 4090 run Llama 3.1 8B Instruct?
Comfortable: GGUF Q8_0 needs 9.5 GB at 4K, batch 1; headroom 14.5 GB.
Text input, FP16 KV cache, one card. Unified-memory capacity uses the GPU guide’s usable limit; negative headroom may require host offload. Speed source settings may differ from this 4K estimate.
Measured runs
| Framework | Quant | tok/s | VRAM used (GB) |
|---|---|---|---|
| ExLlamaV2 | EXL2 4.65bpw | 235 | 5.4 |
| vLLM | AWQ INT4 | 218 | 4.9 |
| llama.cpp | GGUF Q4_K_M | 148 | 5.7 |
Quant memory at 4K
| Quant | Total GB | Verdict | Weight sizing |
|---|---|---|---|
| GGUF Q2_K | 3.3 | Comfortable | Estimated weights |
| GGUF Q4_K_M | 5.6 | Comfortable | Estimated weights |
| GGUF Q6_K | 7.4 | Comfortable | Estimated weights |
| GGUF Q8_0 | 9.5 | Comfortable | Estimated weights |
| AWQ INT4 | 4.8 | Comfortable | Estimated weights |
| EXL2 4.65bpw | 5.4 | Comfortable | Estimated weights |
Context capacity · GGUF Q8_0
| Context | Total GB | Verdict |
|---|---|---|
| 4K | 9.5 | Comfortable |
| 16K | 11.1 | Comfortable |
| 32K | 13.3 | Comfortable |
| 128K | 26.5 | Does not fit |
Generation speed
Measured: 235 tok/s, EXL2 4.65bpw, ExLlamaV2
Measured: 218 tok/s, AWQ INT4, vLLM
Measured: 148 tok/s, GGUF Q4_K_M, llama.cpp
GGUF Q8_0: 83–109 tok/s · Derived range