2 ms·
Semiconductor pricing is largely determined by volume. The Cell processor had mediocre normal performance inline with other Power derivatives but really good ve
by acdha 7y ago
Semiconductor pricing is largely determined by volume. The Cell processor had mediocre normal performance inline with other Power derivatives but really good vector performance for the target price (this was before GPU computing matured), so people with vector-amenable tasks could theoretically win by adopting a processor shipping by the million rather than using lower-volume chips targeted at traditional HPC uses. Since Cell was a Sony/IBM project, IBM was interested in reselling the same work to new customers and (IIRC) needed the volume to have their chip foundry business be healthy.
The problem was similar to Intel’s flop with IA-64: you could get a few amazing benchmarks but code which couldn’t be perfectly scheduled like that performed poorly, and that tuning was hard even if it could be done. The scientists I supported at the time were interested but the numbers made no sense: some code was inherently branch-heavy and couldn’t be tuned much and for the portions which could it was cheaper to buy more regular hardware using the money which would have gone to a high-end programmer, especially when you looked at the long-term maintenance cost of that heavily optimized code (everyone of that era remembered ripping out previous assembly optimizations which had become worse than C on newer CPU & compilers).