{"type":"excalidraw","version":2,"source":"https://AwesomeSystemDesign.dev","elements":[{"angle":0,"strokeColor":"#6577aa","backgroundColor":"#eef1fc","fillStyle":"solid","strokeWidth":1.5,"strokeStyle":"solid","roughness":0.7,"opacity":100,"version":1,"versionNonce":1,"isDeleted":false,"groupIds":["step-1"],"frameId":null,"roundness":null,"boundElements":[{"type":"text","id":"shape-1"}],"updated":0,"link":null,"locked":false,"id":"shape-0","type":"rectangle","seed":41,"x":36,"y":32,"width":341.69296,"height":68,"customData":{"revealStep":1}},{"angle":0,"strokeColor":"#394557","backgroundColor":"transparent","fillStyle":"solid","strokeWidth":1.5,"strokeStyle":"solid","roughness":0.7,"opacity":100,"version":1,"versionNonce":1,"isDeleted":false,"groupIds":["step-1"],"frameId":null,"roundness":null,"boundElements":null,"updated":0,"link":null,"locked":false,"id":"shape-1","type":"text","seed":114,"x":56,"y":54.3,"width":301.69296,"height":23.400000000000002,"fontSize":18,"fontFamily":1,"text":"Precision | Size | Quality | Speed","originalText":"Precision | Size | Quality | Speed","textAlign":"center","verticalAlign":"middle","containerId":"shape-0","lineHeight":1.3,"autoResize":false,"customData":{"revealStep":1}},{"angle":0,"strokeColor":"#628572","backgroundColor":"#edf5ee","fillStyle":"solid","strokeWidth":1.5,"strokeStyle":"solid","roughness":0.7,"opacity":100,"version":1,"versionNonce":1,"isDeleted":false,"groupIds":["step-2"],"frameId":null,"roundness":null,"boundElements":[{"type":"text","id":"shape-3"}],"updated":0,"link":null,"locked":false,"id":"shape-2","type":"rectangle","seed":187,"x":417.6929600000001,"y":32,"width":390,"height":223.20000000000002,"customData":{"revealStep":2}},{"angle":0,"strokeColor":"#394557","backgroundColor":"transparent","fillStyle":"solid","strokeWidth":1.5,"strokeStyle":"solid","roughness":0.7,"opacity":100,"version":1,"versionNonce":1,"isDeleted":false,"groupIds":["step-2"],"frameId":null,"roundness":null,"boundElements":null,"updated":0,"link":null,"locked":false,"id":"shape-3","type":"text","seed":260,"x":437.6929600000001,"y":50,"width":350,"height":187.20000000000002,"fontSize":18,"fontFamily":1,"text":"FP16 (baseline) | 140 GB | 100% |\n1.0×\nINT8 (W8A8) | 70 GB | ~99.5% |\n1.5-2.0×\nINT4 (GPTQ/AWQ) | 35 GB |\n~98-99% | 2.0-3.0×\nFP8 (H100 native)| 70 GB | ~99.8%\n| 1.8-2.0×","originalText":"FP16 (baseline) | 140 GB | 100% |\n1.0×\nINT8 (W8A8) | 70 GB | ~99.5% |\n1.5-2.0×\nINT4 (GPTQ/AWQ) | 35 GB |\n~98-99% | 2.0-3.0×\nFP8 (H100 native)| 70 GB | ~99.8%\n| 1.8-2.0×","textAlign":"center","verticalAlign":"middle","containerId":"shape-2","lineHeight":1.3,"autoResize":false,"customData":{"revealStep":2}},{"angle":0,"strokeColor":"#a68a53","backgroundColor":"#fff7e6","fillStyle":"solid","strokeWidth":1.5,"strokeStyle":"solid","roughness":0.7,"opacity":100,"version":1,"versionNonce":1,"isDeleted":false,"groupIds":["step-3"],"frameId":null,"roundness":null,"boundElements":[{"type":"text","id":"shape-5"}],"updated":0,"link":null,"locked":false,"id":"shape-4","type":"rectangle","seed":333,"x":36,"y":295.20000000000005,"width":637.0026,"height":301.2,"customData":{"revealStep":3}},{"angle":0,"strokeColor":"#394557","backgroundColor":"transparent","fillStyle":"solid","strokeWidth":1.5,"strokeStyle":"solid","roughness":0.7,"opacity":100,"version":1,"versionNonce":1,"isDeleted":false,"groupIds":["step-3"],"frameId":null,"roundness":null,"boundElements":null,"updated":0,"link":null,"locked":false,"id":"shape-5","type":"text","seed":406,"x":56,"y":313.20000000000005,"width":597.0026,"height":265.2,"fontSize":17,"fontFamily":1,"text":"Model Quantization — Reduce Memory & Increase Speed\n\nOriginal Llama-70B in FP16: ~140 GB (too big for 2× 80 GB GPU)\n\nINT4 Llama-70B: 35 GB → fits on a SINGLE 80 GB GPU!\n→ 4× fewer GPUs needed → 4× cost reduction\n→ Quality loss is minimal for most use cases\n\nWhen NOT to quantize:\n• Math-heavy tasks (precision matters)\n• Embedding models (vector quality degrades)\n• When you need exact reproducibility","originalText":"Model Quantization — Reduce Memory & Increase Speed\n\nOriginal Llama-70B in FP16: ~140 GB (too big for 2× 80 GB GPU)\n\nINT4 Llama-70B: 35 GB → fits on a SINGLE 80 GB GPU!\n→ 4× fewer GPUs needed → 4× cost reduction\n→ Quality loss is minimal for most use cases\n\nWhen NOT to quantize:\n• Math-heavy tasks (precision matters)\n• Embedding models (vector quality degrades)\n• When you need exact reproducibility","textAlign":"left","verticalAlign":"middle","containerId":"shape-4","lineHeight":1.3,"autoResize":false,"customData":{"revealStep":3}}],"appState":{"viewBackgroundColor":"#fbfbf8","gridSize":null},"files":{}}