INPUT LEVEL DISPATCH LEVEL COMPRESSION / PACKING EXECUTION PATHS ACCUMULATION & WRITEBACK TERNARY PATH SIMD MAD PATH LUT INDEX SIMD FEED TARGET: RTX 4090 BLOCK: 128 (FIT) KERNEL: JIT TERNARY SHAPE: 4096×4096 MATRIX MULTIPLY DISPATCH NODE PACKED TERNARY LOOKUP TABLE ACCUMULATION PACKED LOW-BIT SIMD MAD BLOCKS ACCUMULATION GPU DECODE 2-BIT PACKED EXPAND / TC TENSOR CORE HIGH-BANDWIDTH MEMORY LANE (UNBLOCKING COMPUTE)