I don't think that would make a very good GPU as it wouldn't fit the data model well [1] as the instructions and data need to share the same bus which would mess up streaming.
Your suggestion is closer to a grid computer but even then I don't think an unmodified 6502 would be a great choice because the memory model (or lack thereof) would really restrict performance.
The LAN controller used to make the Beowulf cluster would probably have more compute (and memory) than the 6502 itself.
The Intel cores in the linked article have a distinct L1 data and instruction caches inside them, and associated L2 caches, which makes a big difference in comparison to the 6502.