{"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":16,"y":16,"width":288,"height":93.2,"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":36,"y":34,"width":248,"height":57.2,"fontSize":22,"fontFamily":1,"text":"Precision | Size |\nQuality | Speed","originalText":"Precision | Size |\nQuality | Speed","textAlign":"left","verticalAlign":"middle","containerId":"shape-0","lineHeight":1.3,"autoResize":false,"customData":{"revealStep":1}},{"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-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":16,"y":153.2,"width":288,"height":350.6,"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":36,"y":171.2,"width":248,"height":314.6,"fontSize":22,"fontFamily":1,"text":"FP16 (baseline) | 140\nGB | 100% | 1.0×\nINT8 (W8A8) | 70\nGB | ~99.5% |\n1.5-2.0×\nINT4 (GPTQ/AWQ) |\n35 GB | ~98-99% |\n2.0-3.0×\nFP8 (H100 native)|\n70 GB | ~99.8% |\n1.8-2.0×","originalText":"FP16 (baseline) | 140\nGB | 100% | 1.0×\nINT8 (W8A8) | 70\nGB | ~99.5% |\n1.5-2.0×\nINT4 (GPTQ/AWQ) |\n35 GB | ~98-99% |\n2.0-3.0×\nFP8 (H100 native)|\n70 GB | ~99.8% |\n1.8-2.0×","textAlign":"left","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":16,"y":547.8,"width":288,"height":865.4,"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":36,"y":565.8,"width":248,"height":829.4,"fontSize":22,"fontFamily":1,"text":"Model Quantization\n— Reduce Memory &\nIncrease Speed\n\nOriginal Llama-70B in\nFP16: ~140 GB (too\nbig for 2× 80 GB\nGPU)\n\nINT4 Llama-70B: 35\nGB → fits on a\nSINGLE 80 GB\nGPU!\n→ 4× fewer GPUs\nneeded → 4× cost\nreduction\n→ Quality loss is\nminimal for most use\ncases\n\nWhen NOT to\nquantize:\n• Math-heavy tasks\n(precision matters)\n• Embedding models\n(vector quality\ndegrades)\n• When you need\nexact reproducibility","originalText":"Model Quantization\n— Reduce Memory &\nIncrease Speed\n\nOriginal Llama-70B in\nFP16: ~140 GB (too\nbig for 2× 80 GB\nGPU)\n\nINT4 Llama-70B: 35\nGB → fits on a\nSINGLE 80 GB\nGPU!\n→ 4× fewer GPUs\nneeded → 4× cost\nreduction\n→ Quality loss is\nminimal for most use\ncases\n\nWhen NOT to\nquantize:\n• Math-heavy tasks\n(precision matters)\n• Embedding models\n(vector quality\ndegrades)\n• When you need\nexact reproducibility","textAlign":"left","verticalAlign":"middle","containerId":"shape-4","lineHeight":1.3,"autoResize":false,"customData":{"revealStep":3}}],"appState":{"viewBackgroundColor":"#fbfbf8","gridSize":null},"files":{}}