{"type":"excalidraw","version":2,"source":"https://AwesomeSystemDesign.dev","elements":[{"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-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":30,"y":30,"width":535.71748,"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":50,"y":52.3,"width":495.71748,"height":23.400000000000002,"fontSize":18,"fontFamily":1,"text":"GPU Costs Dominate — Every Optimization Matters","originalText":"GPU Costs Dominate — Every Optimization Matters","textAlign":"left","verticalAlign":"middle","containerId":"shape-0","lineHeight":1.3,"autoResize":false,"customData":{"revealStep":1}},{"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-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":30,"y":123,"width":506.08624000000003,"height":106.2,"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":50,"y":141,"width":466.08624000000003,"height":70.2,"fontSize":18,"fontFamily":1,"text":"1. Quantization: INT4 → 4× fewer GPUs per model\nLlama-70B: 4× H100 (FP16) → 1× H100 (INT4)\nSavings: ~$72/hr → ~$18/hr per instance","originalText":"1. Quantization: INT4 → 4× fewer GPUs per model\nLlama-70B: 4× H100 (FP16) → 1× H100 (INT4)\nSavings: ~$72/hr → ~$18/hr per instance","textAlign":"left","verticalAlign":"middle","containerId":"shape-2","lineHeight":1.3,"autoResize":false,"customData":{"revealStep":2}},{"angle":0,"strokeColor":"#9180ac","backgroundColor":"#f4effb","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":30,"y":254.2,"width":550.5235600000001,"height":106.2,"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":50,"y":272.2,"width":510.5235600000001,"height":70.2,"fontSize":18,"fontFamily":1,"text":"2. Scale-to-Zero: Don't pay for idle GPUs\n10K endpoints × $18/hr each = $180K/hr if always-on\nWith scale-to-zero (90% idle): ~$18K/hr (10× savings)","originalText":"2. Scale-to-Zero: Don't pay for idle GPUs\n10K endpoints × $18/hr each = $180K/hr if always-on\nWith scale-to-zero (90% idle): ~$18K/hr (10× savings)","textAlign":"left","verticalAlign":"middle","containerId":"shape-4","lineHeight":1.3,"autoResize":false,"customData":{"revealStep":3}},{"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-4"],"frameId":null,"roundness":null,"boundElements":[{"type":"text","id":"shape-7"}],"updated":0,"link":null,"locked":false,"id":"shape-6","type":"rectangle","seed":479,"x":30,"y":385.4,"width":503.01436,"height":129.60000000000002,"customData":{"revealStep":4}},{"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-4"],"frameId":null,"roundness":null,"boundElements":null,"updated":0,"link":null,"locked":false,"id":"shape-7","type":"text","seed":552,"x":50,"y":403.4,"width":463.01436,"height":93.60000000000001,"fontSize":18,"fontFamily":1,"text":"3. Spot/Preemptible Instances for Batch:\nBatch inference: not latency-sensitive\nSpot GPUs: 60-70% cheaper than on-demand\n→ Accept interruption, checkpoint progress, resume","originalText":"3. Spot/Preemptible Instances for Batch:\nBatch inference: not latency-sensitive\nSpot GPUs: 60-70% cheaper than on-demand\n→ Accept interruption, checkpoint progress, resume","textAlign":"left","verticalAlign":"middle","containerId":"shape-6","lineHeight":1.3,"autoResize":false,"customData":{"revealStep":4}},{"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-5"],"frameId":null,"roundness":null,"boundElements":[{"type":"text","id":"shape-9"}],"updated":0,"link":null,"locked":false,"id":"shape-8","type":"rectangle","seed":625,"x":30,"y":540,"width":536.38528,"height":106.2,"customData":{"revealStep":5}},{"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-5"],"frameId":null,"roundness":null,"boundElements":null,"updated":0,"link":null,"locked":false,"id":"shape-9","type":"text","seed":698,"x":50,"y":558,"width":496.38527999999997,"height":70.2,"fontSize":18,"fontFamily":1,"text":"4. Multi-LoRA: 100 models on 1 base\nInstead of 100 full model copies (100× GPUs)\nShare base model, swap 100 MB adapters → 1× GPUs","originalText":"4. Multi-LoRA: 100 models on 1 base\nInstead of 100 full model copies (100× GPUs)\nShare base model, swap 100 MB adapters → 1× GPUs","textAlign":"left","verticalAlign":"middle","containerId":"shape-8","lineHeight":1.3,"autoResize":false,"customData":{"revealStep":5}},{"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-6"],"frameId":null,"roundness":null,"boundElements":[{"type":"text","id":"shape-11"}],"updated":0,"link":null,"locked":false,"id":"shape-10","type":"rectangle","seed":771,"x":30,"y":671.2,"width":557.65948,"height":82.80000000000001,"customData":{"revealStep":6}},{"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-6"],"frameId":null,"roundness":null,"boundElements":null,"updated":0,"link":null,"locked":false,"id":"shape-11","type":"text","seed":844,"x":50,"y":689.2,"width":517.65948,"height":46.800000000000004,"fontSize":18,"fontFamily":1,"text":"5. Continuous Batching: 2-3× throughput per GPU\nSame GPU serves 2-3× more requests → amortize cost","originalText":"5. Continuous Batching: 2-3× throughput per GPU\nSame GPU serves 2-3× more requests → amortize cost","textAlign":"left","verticalAlign":"middle","containerId":"shape-10","lineHeight":1.3,"autoResize":false,"customData":{"revealStep":6}},{"angle":0,"strokeColor":"#9180ac","backgroundColor":"#f4effb","fillStyle":"solid","strokeWidth":1.5,"strokeStyle":"solid","roughness":0.7,"opacity":100,"version":1,"versionNonce":1,"isDeleted":false,"groupIds":["step-7"],"frameId":null,"roundness":null,"boundElements":[{"type":"text","id":"shape-13"}],"updated":0,"link":null,"locked":false,"id":"shape-12","type":"rectangle","seed":917,"x":30,"y":779,"width":638.57776,"height":106.2,"customData":{"revealStep":7}},{"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-7"],"frameId":null,"roundness":null,"boundElements":null,"updated":0,"link":null,"locked":false,"id":"shape-13","type":"text","seed":990,"x":50,"y":797,"width":598.57776,"height":70.2,"fontSize":18,"fontFamily":1,"text":"6. Prefix Caching: Avoid redundant prefill\n500-token system prompt × 20K req/s = 10M redundant tokens\nCache prefix → ~90% prefill savings","originalText":"6. Prefix Caching: Avoid redundant prefill\n500-token system prompt × 20K req/s = 10M redundant tokens\nCache prefix → ~90% prefill savings","textAlign":"left","verticalAlign":"middle","containerId":"shape-12","lineHeight":1.3,"autoResize":false,"customData":{"revealStep":7}},{"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-8"],"frameId":null,"roundness":null,"boundElements":[{"type":"text","id":"shape-15"}],"updated":0,"link":null,"locked":false,"id":"shape-14","type":"rectangle","seed":1063,"x":30,"y":910.2,"width":614.1172,"height":68,"customData":{"revealStep":8}},{"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-8"],"frameId":null,"roundness":null,"boundElements":null,"updated":0,"link":null,"locked":false,"id":"shape-15","type":"text","seed":1136,"x":50,"y":932.5,"width":574.1172,"height":23.400000000000002,"fontSize":18,"fontFamily":1,"text":"Combined savings: 20-50× cost reduction vs naive deployment","originalText":"Combined savings: 20-50× cost reduction vs naive deployment","textAlign":"left","verticalAlign":"middle","containerId":"shape-14","lineHeight":1.3,"autoResize":false,"customData":{"revealStep":8}}],"appState":{"viewBackgroundColor":"#fbfbf8","gridSize":null},"files":{}}