It's not that hard to calculate.
We have 6 SPEs available, each can do one 4*32 bit vector op/clock. clocked at 3.2GHz.
6 * 4 * 3.2G = 76.8GINTOPs
No native bitrotate, no ch()-like opcode, so we need about 6350 ops/bitcoinhash.
76800M / 6350ops/hash = ~ 12.1 Mhash/s
Not THAT bad for a >5 year old design.
About on par with a modern CPU, completely outclassed even by a midrange GPU.
PS3: 12.1Mh/s, ~80W, ~$250
HD5770: 156.8Mh/s, ~110W, ~$130
I liked this analysis, but it turns out to be incorrect. The Cell's SPU does in fact have both a native bit rotate (ROTI) and a Ch()-like opcode (SELB).
Since I'm currently measuring 21.471 Mhash/s with my Cell implementation, I guess you could work backwards and calculate 76800M / 21.471 Mhash/s = ~ 3577ops/hash.
However, I think you're right that it is still outclassed by a midrange GPU.