| bdbdbd said: I believe around 05-06 ARM was largely still sub-200MHz single core and it was roughly at this time that Intel introduced it's Core architecture, so the multicore processors weren't really that common at the time. Technically the Cell could produce high performance per cost, as the SPE's could eventually be dirt cheap to manufacture, opposed to multi-PPE core processors. I don't know if it was because of bad dev tools, as it was the processor that required you to know when an SPE was doing nothing. |
Yes, ARM SoCs were jokes at the time. They only had to power feature phones. Smartphones made them jump in performance quickly, stabilizing after reaching limits imposed by physics. But nowadays, they pack quite a punch.
I remember it was extremely hard to debug stuff on the SPEs. Comparing with CUDA, that has great tools, good docs and helps a lot to deal with the quirks of GPU programming, it's hard to get dev support. Intel is facing this issue now trying to push the Xeon Phi, specially now that tons of libraries support CUDA. But you're correct when pointing that the difficulties were not only on the tools, but also on the weird design of the architecture. It doesn't matter if its ingenuous if it takes way more time to develop something for it.
