An　implementation　of　volumetric　beamforming　for　row-column　addressed　arrays　(RCAs)　is　proposed,　with　optimizations　for　Graphics　Processing　Units　(GPUs).　It　is　hypothesized　that　entire　volumes　can　imaged　in　real　time　by　a　consumer-class　GPU　at　an　emission　rate　≥12　kHz.　A　separable　beamforming　algorithm　was　used　to　reduce　the　number　of　calculations　with　a　negligible　impact　on　the　image　quality.　Here,　a　single　image　was　beamformed　for　each　emission　and　then　extrapolated　to　reproduce　the　volume,　which　resulted　in　65　times　fewer　calculations　per　volume.　Reusing　computations　and　samples　among　adjacent　pixels　and　frames　reduced　the　amount　of　overhead　and　load　instructions,　increasing　performance.　A　GPU　beamformer,　written　in　CUDA　C++,　was　modified　to　implement　the　dual-stage　imaging　with　optimizations.　In-vivo　rat　kidney　data　was　acquired　using　a　6　MHz　Vermon　128+128　RCA　probe　and　a　Verasonics　Vantage　256　scanner.　The　acquisition　used　96　defocused　emissions　at　a　12　kHz　rate　for　a　volume　acquisition　rate　of　125　Hz.　Processing　time,　including　all　pre-processing,　was　measured　for　an　NVIDIA　GeForce　RTX　4090　GPU,　and　the　resulting　beamforming　rate　was　1440　volumes　per　second,　greatly　exceeding　the　real-time　rate.　Based　on　the　GPU＇s　floating-point　throughput,　this　corresponds　to　22%　of　the　theoretically　achievable　rate.　High　efficiency　was　also　shown　for　an　RTX　2080　Ti　and　RTX　3090,　both　achieving　real-time　imaging.　This　shows　that　3D　imaging　can　be　performed　in　real　time　with　a　setup　similar　to　2D　imaging:　Using　a　single　graphics　card,　one　scanner,　and　128　transmit/receive　channels.　　©　1986-2012　IEEE.