TFLOPS vs TOPS

TFLOPS, or teraFLOPS, counts trillions of floating-point operations per second at a stated precision such as FP32 or FP16. TOPS, or tera operations per second, counts trillions of operations without naming a number format, and vendors use it mostly for 8-bit integer (INT8) math. NVIDIA also quotes 4-bit floating point (FP4) throughput in TOPS, under the label AI TOPS. Datasheet figures in both units are peak rates, not measured speed.

The A100 datasheet lists 19.5 TFLOPS at FP32, 312 TFLOPS of FP16 tensor math, and 624 TOPS at INT8, or 1,248 TOPS with sparsity. The RTX PRO 6000 Blackwell Workstation Edition lists 125 TFLOPS of single precision (FP32) beside 4,000 AI TOPS. Its footnote defines that figure as theoretical FP4 with sparsity. The number is 32 times the FP32 rate, counted in a different format. AMD's Ryzen AI Max+ 395 lists up to 60 TFLOPS of FP16, up to 50 neural processing unit (NPU) TOPS and up to 126 overall TOPS. The page names no data type for either TOPS figure.

Compare two cards only on figures that share a data type and a sparsity setting, and read the footnotes to find both. A TOPS figure with no named data type cannot be ranked against any TFLOPS figure.

Sources

Last verified October 7, 2026.