Search Results
7/22/2025, 3:35:03 AM
NEW UDNA GOSSIP (via leaker KeplerL2)
https://www.techspot.com/news/108754-amd-upcoming-rdna-5-flagship-could-target-rtx.html
>Similar to the 7900 XTX, the upcoming high-end AMD GPU will likely include 96 compute units and a 384-bit memory bus. A mid-range version is expected to offer 64 compute units and a 256-bit memory bus, resembling the 9070 XT. A mainstream option might be similar to the 9060 XT, with 32 compute units and a 128-bit bus.
>According to sources familiar with AMD's hardware roadmap, Kepler previously estimated that UDNA will improve raster performance by approximately 20 percent over RDNA 4 and double its ray tracing capabilities. RDNA 4 already represents a significant leap in ray tracing over its predecessor.
AMD could stick with GDDR6(X) to control costs and the only real reason I'd bother with GDDR7 if I were them was for the 3GB VRAM modules to increase the buffer spec without resorting to a clamshell design.
Since this will be a gen in which they get console money to fund it's R&D, the UDNA5 "might" be good, but AMD needs to decide if they care enough to actually mint cards in volume to keep prices competitive and stock plentiful unlike currently against the RTX 5000 series. There is no excuse for what happened with pricing after stockpiling cards for 3 months. Also I am worried with the unified architecture that an obscene amount of die space will be dedicated to worthless compute rendering it deadwood transistors for game performance.
And I've said it before and I'll say it again, whichever team decides to bite the bullet and leave TSMC at the wedding altar and mint their cards on a cheaper lineup (ex: Samsung or Intel's fabs, PPA be damned) and pass some savings to the consumer will win next-gen. AMD might be more inclined since that would free up TSMC allocation to go back into CPUs.
https://www.techspot.com/news/108754-amd-upcoming-rdna-5-flagship-could-target-rtx.html
>Similar to the 7900 XTX, the upcoming high-end AMD GPU will likely include 96 compute units and a 384-bit memory bus. A mid-range version is expected to offer 64 compute units and a 256-bit memory bus, resembling the 9070 XT. A mainstream option might be similar to the 9060 XT, with 32 compute units and a 128-bit bus.
>According to sources familiar with AMD's hardware roadmap, Kepler previously estimated that UDNA will improve raster performance by approximately 20 percent over RDNA 4 and double its ray tracing capabilities. RDNA 4 already represents a significant leap in ray tracing over its predecessor.
AMD could stick with GDDR6(X) to control costs and the only real reason I'd bother with GDDR7 if I were them was for the 3GB VRAM modules to increase the buffer spec without resorting to a clamshell design.
Since this will be a gen in which they get console money to fund it's R&D, the UDNA5 "might" be good, but AMD needs to decide if they care enough to actually mint cards in volume to keep prices competitive and stock plentiful unlike currently against the RTX 5000 series. There is no excuse for what happened with pricing after stockpiling cards for 3 months. Also I am worried with the unified architecture that an obscene amount of die space will be dedicated to worthless compute rendering it deadwood transistors for game performance.
And I've said it before and I'll say it again, whichever team decides to bite the bullet and leave TSMC at the wedding altar and mint their cards on a cheaper lineup (ex: Samsung or Intel's fabs, PPA be damned) and pass some savings to the consumer will win next-gen. AMD might be more inclined since that would free up TSMC allocation to go back into CPUs.
Page 1