Table 9. Scaling Across Multiple GPU Instances^a.

instances	total cores	ranks × threads	MEM (ns/d)	EMEM	RIB (ns/d)	ERIB
1	32	1 × 32/8 × 4	98.1	1.00	7.48	1.00
2	64	8 × 8/32 × 2	76.0	0.39	13.27	0.89
4	128	8 × 16/32 × 4	73.2	0.19	19.50	0.65
8	256	32 × 8			24.39	0.41
16	512	64 × 8			28.38	0.24
32	1024	32 × 32			21.47	0.09

^aAs in Table 7, but for g4dn.16xl instances with hyperthreading off.