Radeon VII wasn’t a lower tier die. Every Radeon VII ever sold was effectively a salvaged Instinct MI50 that was unable to be validated for that particular market segment. It was and remains AMD’s only equivalent to NVIDIA’s (very dead) Titan line of products (as all Titans were salvaged Quadros and Teslas).
The jump from 14nm (which wasn’t really that much different than 16nm) to 7nm can’t be overstated. It was only slightly less of a leap as NVIDIA recently made when jumping from Samsung 8nm to TSMC N4 this generation (which was *massive*). VEGA 20 might be significantly smaller than VEGA 10, but it also packs 10% more transistors into that smaller surface area. Additionally, the memory interface is twice as wide in VEGA 20 (4096 bit) relative to VEGA 10 (2048bit) because AMD doubled the HBM2 stacks from 2 to 4. HBM2 was/is insanely expensive compared to GDDR5, GDDR5x, GDDR6, and GDDR6x modules, so much so that Radeon VII’s VRAM cost *by itself* was equitable to the BOM board partners were paying to manufacture a complete RX 580.
All in all, it was an okay card. It wasn’t particularly good for gaming relative to its peers but the same criticism could easily be made for VEGA 56 and 64. It was an phenomenal buy for content creators who needed gobs of VRAM but couldn’t afford the $2500 NVIDIA was asking for Titan V.
They have their own equivalent in the CDNA line of compute products.
They absolutely could bring matrix multiplication units to their consumer cards, they just refuse to do so.