Here is the benchmarks of AlphaFold (v2.3) I did on ASGC DiCOS system. The performance of Nvidia RTx 3090 (24GB) cards was quite impressed to me, compared to expensive V100 cards. I would conclude to mainly use RTx3090 cards for protein with the total amino acids smaller than 1500 (single polypeptide, oligomer, or complex). For a protein greater than 2000 amino acids, A100 card is highly recommended.

Protein total lengths	A100 80GB	V100 24GB	RTx3090 24GB
112	0.77	1.18	1.07
177	1.07	1.80	1.52
249	1.33	2.85	2.0
336	1.58	4.05	2.85
360	1.48	3.63	2.58
469	2.48	5.33	4.33
936 (homodimer)	6.68	19.33	11.77
1136	10.95	34.57	20.95
1280 (homotetramer)	7.42	22.08	12.75
1872 (homotetramer)	23.82	80.32	44.50
3744 (homooctamer)	153.23	887.93	816.02*

The “rank 001” models calculated by AlphaFold using the 3 different GPU cards are compared one by one. In my observation, there is no deviations in the “structured” regions. If the protein is composed of a portion of intrinsically disordered region (IDR), the overall 3D fold may be slightly vary. If this target is multi-domain protein without a compact conformation, the individual domains are all alike but 3D conformations are different. Here are some examples:

The 3090 card didn’t well handle the case of the octameric protein (3744 aa). All 5 models show similar “dashed, missed” chains, like the example below. The AlphaFold calculation using V100 for the same sequence resulted in correct conformation. No such missed, dashed structure visualized in PyMOL. I am not sure what happened in the AlphaFold calculation, but I will not consider 3090 cards for such large molecular weight proteins. In fact, A100 card is the best candidate for such task as it enormously reduced the computational time.