=== Phase 0 Hardware Microbench v2 (Disjoint Offsets + QD Curve + 60s Thermal) === Disjoint 512 KiB random (std GGUF, F_NOCACHE=1) | th=1 | 1073.7 MB | med 2.27 GB/s (min 1.88, max 2.33) Disjoint 512 KiB random (std GGUF, F_NOCACHE=1) | th=2 | 1073.7 MB | med 5.33 GB/s (min 5.22, max 5.35) Disjoint 512 KiB random (std GGUF, F_NOCACHE=1) | th=4 | 1073.7 MB | med 7.13 GB/s (min 7.13, max 7.13) Disjoint 512 KiB random (std GGUF, F_NOCACHE=1) | th=8 | 1073.7 MB | med 7.13 GB/s (min 7.12, max 7.14) Control: 512 KiB cached warmup (F_NOCACHE=0) | th=4 | 1073.7 MB | med 7.03 GB/s (min 7.03, max 7.03) Control: 512 KiB cached RAM (F_NOCACHE=0) | th=1 | 1073.7 MB | med 25.59 GB/s (min 25.51, max 25.80) Control: 512 KiB cached RAM (F_NOCACHE=0) | th=4 | 1073.7 MB | med 45.70 GB/s (min 45.29, max 45.88) Disjoint 1.5 MiB random (experts' copy, NOCACHE) | th=1 | 1101.0 MB | med 3.39 GB/s (min 2.64, max 3.64) Disjoint 1.5 MiB random (experts' copy, NOCACHE) | th=2 | 1101.0 MB | med 6.70 GB/s (min 4.43, max 6.71) Disjoint 1.5 MiB random (experts' copy, NOCACHE) | th=4 | 1101.0 MB | med 6.83 GB/s (min 3.59, max 6.83) Disjoint 1.5 MiB random (experts' copy, NOCACHE) | th=8 | 1101.0 MB | med 6.77 GB/s (min 6.71, max 6.86) === Metal GPU 60s Sustained Load / Thermal Sag Test (M=4096, K=2560, N=2560 FP16) === t= 0s.. 5s | 1308 GEMMs | 14.00 TFLOP/s t= 5s..10s | 1308 GEMMs | 14.00 TFLOP/s t=10s..15s | 1308 GEMMs | 14.03 TFLOP/s t=15s..20s | 1308 GEMMs | 13.99 TFLOP/s t=20s..25s | 1308 GEMMs | 13.99 TFLOP/s t=25s..30s | 1308 GEMMs | 14.01 TFLOP/s t=30s..35s | 1308 GEMMs | 13.98 TFLOP/s t=35s..40s | 1308 GEMMs | 13.93 TFLOP/s t=40s..45s | 1308 GEMMs | 13.92 TFLOP/s t=45s..50s | 1308 GEMMs | 13.92 TFLOP/s t=50s..55s | 1296 GEMMs | 13.90 TFLOP/s t=55s..60s | 1296 GEMMs | 13.88 TFLOP/s