Is the hack itself not clear enough?
We kinda have a few things down thanks to just looking at the hack itself
- There is a new API for Nintendo made by NVIDIA called NVN2
- NVN2 supports DLSS and RT and has the GPU for an SoC called Drake (Codenamed T239)
- Drake/T239 has a 12SM GPU with 1:1 on SM/RT Cores and is an Ampere-family GPU due to its GA- Designation (GA10F)
- 1536 CUDA Cores
- 12RT cores
- 48 Tensor cores
- 1.5MB L1 (128KB L1 per SM like Desktop Ampere)
- 1 to 4MB L2 Cache
- NVN2 has the hooks for TX1/Pascal GPUs ripped out, meaning it has to be a new SoC on a new Architecture (Further corroborating #3 and that #2 would get used)
- The "last edit" dates for files give us a loose timeline of how development for NVN2 has been and seems to line up with the devkit timeline you yourself mentioned before (NVN2 started being worked on in late 2019, which lines up with first-round devkits going out at the end of 2020)
As for how strong the system will be, that is up to interpretation but based on even just simple TFLOP calculations and the statistic that GCN (PS4/Xbone gen systems) are similar in FLOP-Efficiency to Ampere, assuming no IPC increase for Drake that would mean even at 768MHz the system would be a 2.36TFLOP machine before DLSS and in an optimized setting Ampere should have even more headroom as it has features beyond DLSS/RT that GCN lacks so it's more than just 32% better than PS4 in that case.
And if it has the 4MB L2 that likely would increase the value of a Drake TFLOP versus Ampere and GCN like Infinity Cache does
That is still pretty big stuff at least if it can indeed be validated against your previous info.