Comment by jsheard

The question is whether there's enough overall demand for a GPU architecture with 4x the VRAM of a 5090 but only about 1/3rd of the bandwidth. At that point it would only really be good for AI inferencing, so why not make specialized inferencing silicon instead?