| Darc Requiem said: People can't have it both ways. You can't want AMD to produce a comparable upscaler to Nvidia's, which requires dedicated hardware to run, and expect it to work on architectures without that hardware. Bear in mind, I'm a 7900XTX owner so this affects me. RDNA2 was never going to being to run FSR4 and RDNA3 was a long shot. RDNA3 has AI accelerators but based on the avialable info, their performance lags behind RDNA4. IIRC, @Pemalite you may know this for sure, the AI accelerators for RDNA3 run off the shaders like the Ray Accelerators. They aren't separate dedicated hardware. If I'm correct, that would mean even if they could get it to run on RDNA3 the performance hit may not be worth it. I hope I'm wrong about this, but I wasn't expected a ML upscaler from AMD to run on their older hardware. As for the pricing, the 9070XT pricing is good. I think $550 would have been a homerun but $600 is a good price. The 9070 pricing is another story. It's too close to the XT's price. With the XT at $600, the non-XT should be $500. |
Correct. The A.I Acceleration on RDNA3/2/1 are all done on the shader pipelines.
So whenever you do A.I tasks on those GPU's it takes away from Ray Tracing, it takes away from Rasterization, you are consuming cache, you are consuming resources, you will always have some contention.
nVidia did manage to get around this somewhat with Turing by allowing concurrent FP and INT operations on a CUDA core, but it just makes the core larger and less efficient overall as it's a jack of all trades, master of none with a duplication of some parts of the pipeline, so it's better to just have an optimized core dedicated to the task instead.
AMD Tried the VLIW approach.
All those issues goes away with RDNA4.
There is technically zero reason why FSR4 couldn't be "back-ported" to older RDNA and non-AMD hardware by translating it to bfloat, but the performance hit would not be worth the effort outside of just trying to maintain compatibility.
***************
The pricing of the 9070XT is good, it could have been an extra $50-$100 cheaper, but I am sure that will happen in a couple of months anyway... But we won't get those MSRP prices at retail, not a dog chance in hell.
The vanilla 9070 needed to be $499 to be a real win.
Depending where the 9070XT and 9070 falls in terms of retailer price here in Australia, it will definitely be my next GPU next month to tie me over for the next couple of years.
It's "good enough" from a performance perspective.
AMD wasn't allowed to screw up the 9070XT/9070 launch from a pricing point and they haven't for the most part, now hopefully they can claw back some marketshare and apply pressure to nVidia.
AMD also has the potential for better pricing in the long term...
The Navi 48 for the 9070XT chip will house 53.9 billion transistors in a 356.5 mm^2 die.
Nvidia's GB203 for 5070 Ti contains 45.6 billion transistors in a 378 mm^2 die.
AMD has the smaller and potentially equivalent/faster chip with cheaper RAM and is likely a fair bit cheaper to manufacture compared to the 5070Ti. (Just need real benchmarks to see how the cards fall).
www.youtube.com/@Pemalite
