TPU · Tensor Processing Unit
Google's custom AI accelerator chip, built around a systolic array for matrix math and used across its cloud and models.
Current numbers
192 GBHBM3E per Ironwood (TPU v7) chip @ 7.37 TB/s; 4,614 FP8 TFLOPS; 6x memory vs Trillium
9,216 chipsGoogle Ironwood (TPU v7) scale-up superpod via ICI (9.6 Tb/s/chip) + OCS (48 units, 13,824 ports)
192 GiB / 7.4 TB/sHBM3E per Ironwood TPU v7 (inference-era ASIC); 9,216-chip pods, 42.5 FP8 ExaFLOPS, 4,614 FP8 TFLOPS/chip