Yes, I like your higher CU count better than my estimate. I assumed a lower number of CU to help with yield. But 64 CUs makes much more sense regarding the clocks and power and it's not the mass market SKU. I still think Anaconda will target ~0.5 TF above PS5. That should feel like enough elbow room above PS5 last minute tweak surprises. So I'd guess we'll get something like this:
- Anaconda - 64 CUs @ 1635 - 13.4 TF
I didn't include Lockhart because quite frankly I'm past caring about it and it feels like a real basket case.
If it was down to pure rendering grunt the number of CU could be pared back a lot compared to Anaconda because 1080 is only 25% of the pixel count of 4k. The number of CUs can't be reduced purely based upon rendering output alone because CU availability must be taken into account for compute tasks (AI, physics, path finding etc).
More and more tasks will be run on CUs in the next gen because CUs will be more abundant. An Anaconda title fully utilising CUs for both rendering and compute in an engine is going to be very tricky to run on Lockhart at the same performance level, even with the smaller render area. You can't just skip engine essential compute tasks if the game relies upon them. Having CPU fall backs seems like ridiculous overkill and will really hamper development work. Titles at the start of the generation probably won't suffer from this as the full hardware spec will be under utilised. At the end of the generation where every cycle is being wrung from the hardware this will definitely cause some issues.
For rendering only 26 CUs should be plenty for rendering at 1080 (assuming the same clock). But the actual number will probably be this plus some number plucked out of thin air to give the future amount of elbow room that more demanding Anaconda compute based titles will need. I wouldn't be surprised to see 30 CUs or even more.
I'm still think we're looking at a Vega base with additions from where ever it was decided to pluck them from, including Navi.