{"model": "reduced-order LLM quantization model -- uniform-quant SQNR accuracy proxy + memory footprint + decode roofline (educational)", "inputs": {"model_params_b": 70.0, "weight_bits": 8.0, "activation_bits": 16.0, "kv_bits": 16.0, "group_size": 128, "batch_size": 32, "seq_length": 4096, "hidden_dim": 8192, "num_layers": 80, "hbm_bandwidth_gb_s": 3350.0, "compute_tops_fp16": 990.0}, "outputs": {"format": "W8A16 KV16", "effective_weight_bits": 8.125, "scale_overhead_percent": 1.54, "weight_gb": 71.094, "kv_cache_gb": 343.597, "total_gb": 414.691, "fp16_total_gb": 483.614, "memory_reduction_x": 1.17, "decode_tokens_s": 258.5, "fp16_decode_tokens_s": 221.7, "decode_speedup_x": 1.17, "compute_speedup_x": 1.0, "decode_bottleneck": "memory bandwidth", "sqnr_weight_db": 37.9, "sqnr_activation_db": 82.5, "sqnr_kv_db": 86.0, "min_sqnr_db": 37.9, "quant_error_percent": 1.28, "accuracy_retention_percent": 100.0, "error_source": "weights", "verdict": "near-lossless -- W8A16 KV16 shrinks the footprint 1.2x and lifts decode throughput 1.2x while retaining 100.0% of full-precision quality. This is a safe production setting; decode is memory bandwidth-bound so the bandwidth savings translate almost directly into tokens/s"}, "profile": {"acc_vs_bits": [[2.0, 67.02], [2.3, 76.73], [2.6, 83.93], [2.89, 89.05], [3.19, 92.61], [3.49, 95.05], [3.79, 96.7], [4.09, 97.8], [4.38, 98.54], [4.68, 99.03], [4.98, 99.36], [5.28, 99.57], [5.57, 99.72], [5.87, 99.81], [6.17, 99.88], [6.47, 99.92], [6.77, 99.95], [7.06, 99.96], [7.36, 99.98], [7.66, 99.98], [7.96, 99.99], [8.26, 99.99], [8.55, 100.0], [8.85, 100.0], [9.15, 100.0], [9.45, 100.0], [9.74, 100.0], [10.04, 100.0], [10.34, 100.0], [10.64, 100.0], [10.94, 100.0], [11.23, 100.0], [11.53, 100.0], [11.83, 100.0], [12.13, 100.0], [12.43, 100.0], [12.72, 100.0], [13.02, 100.0], [13.32, 100.0], [13.62, 100.0], [13.91, 100.0], [14.21, 100.0], [14.51, 100.0], [14.81, 100.0], [15.11, 100.0], [15.4, 100.0], [15.7, 100.0], [16.0, 100.0]], "tps_vs_bits": [[2.0, 296.0], [2.3, 293.9], [2.6, 291.8], [2.89, 289.7], [3.19, 287.7], [3.49, 285.7], [3.79, 283.7], [4.09, 281.8], [4.38, 279.9], [4.68, 278.0], [4.98, 276.1], [5.28, 274.3], [5.57, 272.4], [5.87, 270.7], [6.17, 268.9], [6.47, 267.1], [6.77, 265.4], [7.06, 263.7], [7.36, 262.0], [7.66, 260.4], [7.96, 258.7], [8.26, 257.1], [8.55, 255.5], [8.85, 253.9], [9.15, 252.4], [9.45, 250.8], [9.74, 249.3], [10.04, 247.8], [10.34, 246.3], [10.64, 244.9], [10.94, 243.4], [11.23, 242.0], [11.53, 240.6], [11.83, 239.2], [12.13, 237.8], [12.43, 236.4], [12.72, 235.1], [13.02, 233.7], [13.32, 232.4], [13.62, 231.1], [13.91, 229.8], [14.21, 228.5], [14.51, 227.3], [14.81, 226.0], [15.11, 224.8], [15.4, 223.6], [15.7, 222.4], [16.0, 221.2]], "weight_gb": 71.094, "kv_gb": 343.597, "fp16_total_gb": 483.614, "total_gb": 414.691, "retention_percent": 100.0, "decode_tokens_s": 258.5, "weight_bits": 8.0, "group_size": 128}, "steps": 96, "node": "ip-172-26-4-148", "compute_ms": 0.4}