Performance Board

2023 Class

cDSP of SM8550P SoC

Table 1 Sustained performance on the compute DSP of the SM8550P SoC for the qfloat32 (FP32 input and output) matrix kernel C+=AB with M=192, N=4, K=128, ldA=192, ldB=128, ldC=128. A theoretical peak of 48 GFLOPS is assumed. The performance is given for kernels using HVX vector instructions.

Team	Time (s)	#executions	GFLOPS	%peak
alex	0.572	100000	34.36	71.6

Graviton3 (c7g.xlarge)

Table 2 Sustained performance on the Graviton3 processor for the single precision matrix kernel C+=AB with M=16, N=6, K=1, ldA=16, ldB=1, ldC=16. A theoretical peak of 64 GFLOPS is assumed. The performance is given for kernels using ASIMD vector instructions.

Team	Time (s)	#executions	GFLOPS	%peak
GEMMbler	1.101	120000000	20.92	32.7
peak climber	1.034	100000000	18.56	29.0
Alex’s ASM	1.287	100000000	14.92	23.3
LIBXSMM, 59410c81 (ASIMD)	1.811	100000000	10.60	16.6

Table 3 Sustained performance on the Graviton3 processor for the single precision matrix kernel C+= AB with M=16, N=6, K=48, ldA=16, ldB=48, ldC=16. A theoretical peak of 64 GFLOPS is assumed. The performance is given for kernels using ASIMD vector instructions.

Team	Time (s)	#executions	GFLOPS	%peak
GEMMbler	1.69	10000000	54.24	84.8
peak climber	17.95	100000000	51.35	80.2
Alex’s ASM	18.08	100000000	50.97	79.6
LIBXSMM, 59410c81 (ASIMD)	18.74	100000000	49.17	76.8

Table 4 Sustained performance on the Graviton3 processor for the single precision matrix kernel C+=AB with M=32, N=6, K=1, ldA=32, ldB=1, ldC=32. A theoretical peak of 64 GFLOPS is assumed. The performance is given for kernels using SVE vector instructions.

Team	Time (s)	#executions	GFLOPS	%peak
LIBXSMM, 3253da84 (SVE)	1.883	100000000	20.40	31.9
GEMMbler	1.004	50000000	19.13	29.9
Alex’s ASM	1.198	50000000	16.02	25.0

Table 5 Sustained performance on the Graviton3 processor for the single precision matrix kernel C+= AB with M=32, N=6, K=48, ldA=32, ldB=48, ldC=32. A theoretical peak of 64 GFLOPS is assumed. The performance is given for kernels using SVE vector instructions.

Team	Time (s)	#executions	GFLOPS	%peak
LIBXSMM, 3253da84 (SVE)	3.106	10000000	59.34	92.7
Alex’s ASM	1.572	5000000	58.61	91.6

Table 6 Sustained performance on the Graviton3 processor for the single precision matrix kernel C+= AB with M=128, N=6, K=48, ldA=128, ldB=48, ldC=128. A theoretical peak of 64 GFLOPS is assumed. The performance is given for kernels using SVE vector instructions.

Team	Time (s)	#executions	GFLOPS	%peak
LIBXSMM, 3253da84 (SVE)	12.293	10000000	59.97	93.7
GEMMbler	6.160	5000000	59.84	93.5
Alex’s ASM	2.507	2000000	58.82	91.9

Table 7 Sustained performance on the Graviton3 processor for the single precision matrix kernel C+= AB with M=128, N=48, K=48, ldA=128, ldB=48, ldC=128. A theoretical peak of 64 GFLOPS is assumed. The performance is given for kernels using SVE vector instructions.

Team	Time (s)	#executions	GFLOPS	%peak
LIBXSMM, 3253da84 (SVE)	9.887	1000000	59.66	93.2
GEMMbler	1.486	150000	59.55	93.0
Alex’s ASM	2.039	200000	57.87	90.4

Table 8 Sustained performance on the Graviton3 processor for the bfloat16 matrix kernel C+= AB with M=16, N=12, K=4 with BFMMLA-tailored data layout. A theoretical peak of 256 GFLOPS is assumed. The performance is given for kernels using SVE vector instructions.

Team	Time (s)	#executions	GFLOPS	%peak
GEMMbler	2.056	100000000	74.72	29.2
peak climber	1.464	50000000	52.45	20.5

Table 9 Sustained performance on the Graviton3 processor for the bfloat16 matrix kernel C+= AB with M=16, N=12, K=48 with BFMMLA-tailored data layout. A theoretical peak of 256 GFLOPS is assumed. The performance is given for kernels using SVE vector instructions.

Team	Time (s)	#executions	GFLOPS	%peak
peak climber	4.004	50000000	230.19	89.9
GEMMbler	1.839	20000000	200.49	78.3

2022 Class: A64FX

Table 10 Sustained performance on the A64FX processor for the single precision matrix kernel C+=AB with M=64, N=6, K=1, ldA=64, ldB=1, ldC=64.

Team	Time (s)	#executions	GFLOPS	%peak
LIBXSMM, cdf74576	2.963	50000000	12.96	11.25
HPC-Lovers	3.727	50000000	10.30	8.94
😎	3.753	50000000	10.23	8.88
😎	0.761	10000000	10.09	8.76
Alex’s ASM	3.819	50000000	10.05	8.72

Table 11 Sustained performance on the A64FX processor for the single precision matrix kernel C+=AB with M=64, N=6, K=48, ldA=64, ldB=48, ldC=64.

Team	Time (s)	#executions	GFLOPS	%peak
HPC-Lovers	2.167	5000000	85.88	74.55
LIBXSMM, cdf74576	2.174	5000000	84.78	73.60
Alex’s ASM	2.264	5000000	81.43	70.68
😎	4.547	10000000	81.08	70.38
😎	2.438	5000000	75.61	65.64

Table 12 Sustained performance on the A64FX processor for the single precision matrix kernel C+=AB with M=128, N=6, K=48, ldA=128, ldB=48, ldC=128.

Team	Time (s)	#executions	GFLOPS	%peak
LIBXSMM, 21a5c464	1.721	2000000	85.70	74.39
Alex’s ASM	1.792	2000000	82.30	71.44
😎	8.963	10000000	82.26	71.40
HPC-Lovers	0.901	1000000	81.82	71.02
😎	1.823	2000000	80.89	70.21

Table 13 Sustained performance on the A64FX processor for the single precision matrix kernel C+=AB with M=128, N=48, K=48, ldA=128, ldB=48, ldC=128.

Team	Time (s)	#executions	GFLOPS	%peak
LIBXSMM, 21a5c464	1.437	200000	82.01	71.13
HPC-Lovers	1.453	200000	81.16	70.45
Alex’s ASM	1.484	200000	79.52	69.03

Table 14 Sustained performance on the A64FX processor for the single precision matrix kernel C+=AB with M=63, N=6, K=48, ldA=63, ldB=48, ldC=63.

Team	Time (s)	#executions	GFLOPS	%peak
😎	4.523	10000000	80.22	69.64
HPC-Lovers	0.910	2000000	79.68	69.16

2021 Class: Graviton2

Table 15 Sustained performance on the Graviton2 processor for the single precision matrix kernel C+=AB with M=16, N=4, K=4, ldA=16, ldB=4, ldC=16.

Team	Time (s)	#executions	GFLOPS	%peak
Antonio	2.001	100000000	25.58	64.0
Markus’ ASM	2.101	100000000	24.36	60.9
LIBXSMM, 6c389dbc	2.446	100000000	20.93	52.3
Felix	2.545	100000000	20.12	50.3
Alex’s ASM	2.570	100000000	19.92	49.8

Table 16 Sustained performance on the Graviton2 processor for the single precision matrix kernel C+= AB with M=16, N=4, K=12, ldA=16, ldB=12, ldC=16.

Team	Time (s)	#executions	GFLOPS	%peak
Antonio	4.563	100000000	33.66	84.2
Markus’ ASM	4.603	100000000	33.37	83.4
LIBXSMM, 6c389dbc	4.989	100000000	30.78	77.0
Felix	5.136	100000000	29.91	74.8
Alex’s ASM	5.173	100000000	29.69	74.2

Table 17 Sustained performance on the Graviton2 processor for the single precision matrix kernel C+=AB with M=19, N=4, K=4, ldA=19, ldB=4, ldC=19.

Team	Time (s)	#executions	GFLOPS	%peak
Antonio	3.042	100000000	19.99	50.0
Markus	3.161	100000000	19.23	48.1
Alex’s ASM	3.761	100000000	16.17	40.4
Gipfelstürmer	3.78	100000000	16.04	40.1
LIBXSMM, c4068710	4.013	100000000	15.15	37.9

Table 18 Sustained performance on the Graviton2 processor for the single precision matrix kernel C+=AB with M=32, N=32, K=32, ldA=32, ldB=32, ldC=32.

Team	Time (s)	#executions	GFLOPS	%peak
Markus	1.739	1000000	37.67	94.1
Antonio	1.743	1000000	37.61	94.0
Gipfelstürmer	1.74	1000000	37.54	93.8
Alex’s ASM	1.747	1000000	37.52	93.8
LIBXSMM, c4068710	1.837	1000000	35.67	89.2

Table 19 Sustained performance on the Graviton2 processor using Just-In-Time code generation. Given is the mean performance of C+=AB for the four configs M=x, N=4, K=4, ldA=x, ldB=4, ldC=x where x is in {16, 17, 18, 19}.

Team	Time (s)	#executions	GFLOPS	%peak
Antonio	2.585	100000000	22.05	55.1
Alex’s mini_jit	N/A	100000000	20.16	50.4
LIBXSMM, 5fd40afe	3.345	100000000	17.11	42.8