I gave this a try myself, just for fun, and only managed to obtain satisfying results with hardware sprites, and no alpha :unsure:
What's most surprising is that real-time circles with alpha are a lot faster than pre-rendered sprites with alpha!
With 60 circles and alpha, my frame rate drops to the low 20's every time there's a significant number of explosions. My code is very simple, and all slowdowns happen due to the alpha blending. I thought Fenix was based on SDL, so how come it is faster with alpha blending than C with SDL? :-/