| context_tokens,cache_type,kv_buffer_mib,gpu_mem_mib,prompt_tps,gen_tps,notes |
| 0,f16,768,23092,,,"KV buffer from llama.cpp verbose, baseline GPU from nvidia-smi" |
| 1493,f16,768,23114,923.2,45.2, |
| 5916,f16,768,23116,1210.8,44.7, |
| 11814,f16,768,23116,1187.8,44.9, |
| 23610,f16,768,23116,1153.4,44.6, |
| 110019,f16,768,23116,815.0,38.0,64K+ context |
| 0,q8_0,408,22732,,,"47% smaller KV buffer vs f16" |
| 1493,q8_0,408,22754,925.1,45.3, |
| 5916,q8_0,408,22756,1206.7,44.9, |
| 11814,q8_0,408,22762,1183.9,42.9, |
| 23610,q8_0,408,22774,1149.1,39.7, |
| 110019,q8_0,408,22856,810.2,25.0,64K+ context — gen speed degraded 34% |
| 0,q4_0,216,22540,,,"72% smaller KV buffer vs f16" |
| 1493,q4_0,216,22562,925.7,45.6, |
| 5916,q4_0,216,22564,1205.8,45.0, |
| 11814,q4_0,216,22570,1191.2,42.7, |
| 23610,q4_0,216,22582,1151.9,39.3, |
| 110019,q4_0,216,22664,812.6,24.0,64K+ context — gen speed degraded 37% |
|
|
