Large discrepancy between DSE-predicted and on-board GOPS for 32K-bit multiplication

[Prathmesh-shalu141]

Hi @ZhuopingYang ,

I ran the 32K-bit (32768-bit) multiplication design on the Versal board using 2 PEs.

The AIM DSE predicts a performance of approximately 1920 GOPS for 2 PEs. However, the measured on-board performance is around 654K inst/sec.

Using the following formula to convert the measured inst/sec to GOPS:

inst/sec = (batch × repeat × PE) / total_time

GOPS = (inst/sec × 2 × [N/32]^2) / 10^9

where N = 32768 bits, the measured performance is:

GOPS = (654,000 × 2 × [32768/32]^2) / 10^9
≈ 1371.6 GOPS

Therefore, the difference between the DSE prediction and the measured performance is:

Error = |1920 - 1371.6| / 1920 × 100
≈ 28.56%

This is significantly higher than the reported maximum DSE prediction error of approximately 6%.

I am also attaching the formulas used to calculate inst/sec and GOPS for verification. Could you please check whether these formulas are correct and let us know if there is any additional factor that should be considered when comparing the DSE prediction with the on-board measurement?

Thanks.

[ZhuopingYang]

Please check the frequency settings in the DSE model. The frequency used in the model is usually larger than the achieved frequency, and the model doesn't predict the final achievable frequency.

[Prathmesh-shalu141]

@ZhuopingYang in AIM.conf I have kept freq=200 and in the freq.txt file, which is in the template folder, it also shows 200 only. Is there any other way to check on-board achieved frequency?

Additional :
I have ran DSE for various bitwidth and pl freq for ex - 8k with 200,160,120 etc freq and in both case dse always gives same gops.

cselab@cselab-Precision-3660:~/AIM-main/AIM/templete$ python3 tools/AIM.py
Your config is as follows:

{'Config': {'input_bitwidth': '8192', 'pl_freq': '200'}, 'Constraints': {'luts': '899840', 'aie_cols': '50', 'aie_rows': '8', 'plio_in': '156', 'plio_out': '117', 'max_row': '19', 'max_col': '20'}, 'Profiling': {'ddr_bandwidth': '15', 'pl_reorganize': '18', 'pl_receive': '10', 'merge_phase2': '11', 'split_input_a': '188', 'split_input_b': '486', 'send_plio': '205', 'receive_plio': '8781', 'serialize_merge': '303', 'non_repeat_luts': '32440'}}

The optimized configuration has been found!
Generating customized AIM architecture...
Your customized AIM architecture has been generated!
The expected performance is 480.00 GOPS
PE: 6
Optimal Row (r): 4
Optimal Column (c): 5
Total AIEs used: 120

cselab@cselab-Precision-3660:~/AIM-main/AIM/templete$ python3 tools/AIM.py
Your config is as follows:

{'Config': {'input_bitwidth': '8192', 'pl_freq': '120'}, 'Constraints': {'luts': '899840', 'aie_cols': '50', 'aie_rows': '8', 'plio_in': '156', 'plio_out': '117', 'max_row': '19', 'max_col': '20'}, 'Profiling': {'ddr_bandwidth': '15', 'pl_reorganize': '18', 'pl_receive': '10', 'merge_phase2': '11', 'split_input_a': '188', 'split_input_b': '486', 'send_plio': '205', 'receive_plio': '8781', 'serialize_merge': '303', 'non_repeat_luts': '32440'}}

The optimized configuration has been found!
Generating customized AIM architecture...
Your customized AIM architecture has been generated!
The expected performance is 480.00 GOPS
PE: 6
Optimal Row (r): 4
Optimal Column (c): 5
Total AIEs used: 120
cselab@cselab-Precision-3660:~/AIM-main/AIM/templete$

[Prathmesh-shalu141]

@ZhuopingYang, please reply.