Cyran said:
A little correction while it true RDNA 2 did not have anything like Tensor correction, RDNA 3 did add in Matrix cores which is AMD version of Tensor cores. Both 20 series and 30 series for Nvidia tensor cores did not support FP8 yet they still give the option to use DLSS 4.5. There is a trade off about a 30% performance drop since it don't have native FP8 support. There no reason AMD could not given RDNA 3 users the option for that trade off like Nvidia did as the nvida 20/30 series was in the same position RDNA 3 was hardware wise. |
I will have to correct your correction.
RDNA 3 did not have dedicated Matrix "cores" like CDNA in AMD's Instinct line, AMD included Matrix Accelerators within their Compute Units and could work alongside the vector units, essentially borrowing the resources that would be used for the FP16 units.
There is still the issue of contention in the CU units due to schedulers and cache being needed for other tasks, it's a shared resource approach.
However, it still doesn't support FP8... And that's the big issue.
RDNA 2 could also do Matrix multiplications in it's vector units in order to support Machine Learning.
...And that is partly the issue here, RDNA1 (Couldn't do it at all), RDNA2, RDNA3, RDNA4 handles Matrix all differently, it's AMD's own fault in that regard, but RDNA4 has aligned itself with the general Industry trend.
And I agree, that AMD could support FSR4 on older Radeon architectures by bifurcating the upscaler, but if we are talking what's best for the consumer, then the best approach for the consumer would be for AMD to make FSR4 open source and let the community manage it and they can just focus all their resources on FSR5.
In the end though, I have never bought a GPU with the idea I am going to get some "new" feature years down the line, it's simply unrealistic, you judge the hardware how it presents on release and at the time of purchase.
I.E. When S3 promised to provide TnL on it's Savage and never did as the TnL unit was buggy.
www.youtube.com/@Pemalite
