Sure, you're not mandated, but then you're going to stagnate even further since research and development won't be done for free, and slashing R&D budget ultimately means slower progress and/or a worse resulting product. Given the reliance on compute modern games have, and that dedicated graphics cards are set to become the primary accelerator in the modern AI PC, reunifying CDNA and RDNA into UDNA, while also lowering their development and maintenance costs and reining in R&D spending is a winning move.
I don't think their previous woes had anything to do specifically with them an unified architecture, but because of the resulting product. At the end of the day, that is what most customers care about. The Vega series (especially the VII) were built on what was essentially an almost 10 year old architectural base, even though they were refined to the extreme and used highly advanced technology for their time (such as HBM memory), the fundamental problems of GCN still affected them, all the while Turing was already DX12 Ultimate capable before the standard was even made public. RDNA 1 was an attempt to restart the gaming lineup, and while the first iteration was pretty bad, we can see that RDNA 2 was very much a respectable architecture, arguably the best AMD's made in the past few years.
It was clearly on a level all its own, even though the prices were far higher than what AMD practiced at the time, you usually got what you paid for, perhaps it didn't mean much at the time with rudimentary, almost demo-like support for DLSS 1 and early RT games in the beginning, but long-term, I'd say people who invested in Turing ended up way better off. We see a similar situation now; Nvidia has a more refined, feature-complete, efficient product to sell, and even though they charge for the privilege, the market has clearly - and repeatedly chosen to support that path. The overwhelming success of Pascal architecture and the fact that people are still fond of and using their GTX 1080 and 1080 Ti GPUs today, 8 years later, has even prolonged the RDNA 1 cards' viability, since they have incomplete DirectX 12 support just like Pascal cards and game developers are still inclined to support these products today.
