id	author	title	date	pages	extension	mime	words	sentences	flesch	summary	cache	txt
work_j6zirmtczrav3e56ilniy62vbi	Massimiliano Fasi	Numerical behavior of NVIDIA tensor cores	2021	19	.pdf	application/pdf	10811	1223	64	We explore the floating-point arithmetic implemented in the NVIDIA tensor cores, � Is the result of each floating-point operation in (2) normalized, or do tensor cores only available, in order to guarantee reproducibility and facilitate testing other matrix multiplyaccumulate units, such as the third generation tensor cores in the latest NVIDIA A100 Rounding modes in tensor core computations Tests for determining what rounding modes are used in the inner products and the final rounding Features of the accumulator Tests that explore the number of extra bits in the alignment step of floating-point addition inside 2. Can tensor cores take binary32 subnormal numbers as inputs for C in (2) without 2. Can tensor cores take binary32 subnormal numbers as inputs for C in (2) without 3. Can tensor cores compute subnormal numbers from normal numbers and return them? When tensor cores are used in binary16 mode, the result computed in the format of	./cache/work_j6zirmtczrav3e56ilniy62vbi.pdf	./txt/work_j6zirmtczrav3e56ilniy62vbi.txt