Link.
I'm really not particularly surprised. It was possible to make a case for the Cell when GPUs were very inefficient at double-precision floating point math, but that has changed a lot recently. Especially with nVidia's forthcoming chip taking on L2 cache and ECC memory, not to mention it's efficiency at 64-bit FP.
It was a cool idea, but it was too late and too slow (both clockspeed and performance).
