Audit weight memory, KV cache, and runtime overhead before treating unified memory as a no-quantization guarantee.
This audit uses binary gibibytes. For seventy billion parameters, two bytes per FP16 parameter require about one hundred thirty point three eight five gibibytes. Half a byte for Q4 requires about thirty-two point five nine six. Both bars use three point five pixels per gibibyte and the same parameter count. The source divides bytes by two to the thirtieth power. This compares assigned weight storage, not output quality or measured allocator usage. Cache storage multiplies two for keys and values, eighty layers, context tokens, eight KV heads, one hundred twenty-eight head dimensions, and two bytes per value. After dividing by two to the thirtieth, eight thousand one hundred ninety-two tokens require two point five gibibytes; doubling context requires five. Both bars use eighty-five pixels per gibibyte. Weight precision does not alter this cache term because cache precision is a separate entered input. At default context, Q4 weights plus cache total about thirty-five point zero nine six gibibytes. This audit applies ten percent overhead to the entire base, including cache, giving about thirty-eight point six zero six gibibytes. Both bars use twelve pixels per gibibyte. The resulting margin against ninety-six gibibytes is about fifty-seven point three nine four. The FP16 scenario exceeds capacity. Unlike the related decimal planner, this source uses binary units and overhead on both weights and cache. Preserve that distinction.