FP4

• Blackwell (microarchitecture), Nvidia claims 20 petaflops of FP4 compute with the dual GB100 die B100 accelerator

• AGX Thor developer kit 2,070 petaflops

See also

• Blackwell
• Floating point, FP4, FP6, FP8, FP16, FP32, FP64, bc, Teraflop, INT8, bfloat16 floating-point format (bf16), mpdecimal