GEMM FPGA Benchmark Results

The benchmark results are given divided by the used version of the benchmark, since internal changes in the benchmark code might lead to different performance results. All measurements were done with single precision floating point matrices of size 4096x4096 which equals 64 MB of data. If this size was not evenly dividable by the number of replications, the matrix size was further reduced to achieve equal load for every kernel replication. The measurements were executed 10 times and the best result is published.

The results and the used configuration is given in Table 8 and are also available as CSV.

Table 8 GEMM FPGA Benchmark Results

Version	1.4	1.0	1.0	1.0	1.0
FPGA board	Bittware 520N	BittWare 520N	Alveo U280	Alveo U280	PAC D5005
FPGA	Intel Stratix 10 GX2800	Intel Stratix 10 GX2800	Xilinx XCU280	Xilinx XCU280	Intel Stratix 10 SX
Memory Type	DDR	DDR	DDR	HBM2	SVM
SDK	21.2.0	19.4.0	2019.2	2019.2	19.4.0
BSP/Shell	20.4.0_hpc	19.2.0_hpc	2019.2.3	2019.2.3	18.1.2_svm
CPU	AMD EPYC Milan 7763	Intel Xeon Gold 6148	Intel Xeon Gold 6148	Intel Xeon Gold 6148	Intel Xeon Gold 6148
System	Noctua 2	Noctua 1	Noctua 1	Noctua 1
BLOCK_SIZE	512	512	256	256	512
GEMM_SIZE	8	8	8	8	8
GLOBAL_MEM_UNROLL	8	16	16	16	16
DATA_TYPE	float	float	float	float	float
NUM_REPLICATIONS	5	5	3	3	5
LUT	310564	275754	568558	499002	299427
LUT percent	33	36.0	51.87	42.64	33.0
Register	793535	861277	441602	920127	829802
Register percent		36.0	19.43	38.7	33.0
BRAM	8321	8860	666	666	9041
BRAM percent	71	76.0	43.11	36.71	77.0
DSP	3318	3398	7683	7683	3398
DSP percent	58	59.0	85.23	85.18	59.0
Frequency	273.07	160.42	100.00	236.00	225.00
GFLOPs	1232.50	708.95	266.91	603.86	739.59
GFLOPs norm	90.27	88.39	85.29	88.97	65.74
Error	9.15527e-5	6.0e-7	2.0e-6	2.0e-6	6.0e-7