A Low-Power Scalable Stream Compute Accelerator for General Matrix Multiply (GEMM)

Table 3

FPGA accelerator performance versus high-end quad core PC.

Matrix size	Ops; (MMACs)	PC time; (s)	FPGA compute time (s)	FPGA system; cycles	FPGA core; cycles	System to core; ratio	PC perf; (GMACS)	FPGA Perf; (GMACS)	PC versus FPGA	FPGA perf at 1 : 2 rat.	FPGA versus PC; at 1 : 2 rat.
4096×4096	68,700	5.25	12.7	2.53 G	491 M	5.15	13.1	5.43	2.4×	56	𝟒.𝟑×
2048×2048	8,500	0.704	1.65	328 M	63.9 M	5.13	12.1	5.22	2.3×	54	4.5×
1024×1024	1,070	0.120	0.221	44 M	8.62 M	5.10	8.91	4.86	1.8×	50	5.7×
512×512	134	0.0235	0.0383	7.6 M	1.37 M	5.54	5.70	3.50	1.62×	39	𝟔.𝟖×
128×128	2.1	0.00081	0.00207	415 k	59 k	7.03	2.59	1.01	𝟐.𝟓𝟔 ×	14	5.5×
4080×1024	17,045	1.403	2.629	526 M	105 M	5.00	12.1	6.48	1.87×	65	5.3×
204×1024	42.6	0.0076	0.0121	460 k	2.4 M	5.21	5.58	3.5	𝟏.𝟓𝟗×	36	6.6×

The bold font emphasizes important entries for comparison reasons.