Jowi Morales Social Links Navigation Contributing Writer Jowi Morales is a tech enthusiast with years of experience working in the industry. He’s been writing with several tech publications since 2021, where he’s been interested in tech hardware and consumer electronics.
VizzieTheViz I’d quite like this for my 7900xtx to be available eventually, but I’ve never had to use any version of fsr yet to get games playable. Not going to bother with all kind of workarounds to get it at any rate. It’d be decent of AMD to release it for rdna3 at some point, even when it’s only when they’ve got version 5 out. Reply
mitch074 RDNA4 has native support for stacked 8-bit floating point instructions, which are needed for fast FSR4 support. RDNA3 does not, but AMD's erroneous commit contained an 8-bit integer version that worked quite well with RDNA3 – the difference in data type means that there are rounding errors that cause the RDNA3 version to look a bit worse. RDNA2 and older don't support WMMA natively, which requires further emulation (also provided by AMD) with a higher cost in performance and even more rounding errors making FSR4 much slower and more buggy on those older cards. So, I'm not surprised AMD is keeping them non-official for now : they get enough backlash from buggy features without adding some more because of even more buggy backports. I'm more surprised AMD isn't making a fuss about these features entering mainline Mesa and Linux kernel when it's now the driver branch they officially support. At the rate it's going, I wonder how long it's going to take for FSR2+ to be backported to even older GCN, like RT before it… Reply
qwertymac93 mitch074 said: RDNA4 has native support for stacked 8-bit floating point instructions, which are needed for fast FSR4 support. RDNA3 does not, but AMD's erroneous commit contained an 8-bit integer version that worked quite well with RDNA3 – the difference in data type means that there are rounding errors that cause the RDNA3 version to look a bit worse. RDNA2 and older don't support WMMA natively, which requires further emulation (also provided by AMD) with a higher cost in performance and even more rounding errors making FSR4 much slower and more buggy on those older cards. So, I'm not surprised AMD is keeping them non-official for now : they get enough backlash from buggy features without adding some more because of even more buggy backports. I'm more surprised AMD isn't making a fuss about these features entering mainline Mesa and Linux kernel when it's now the driver branch they officially support. At the rate it's going, I wonder how long it's going to take for FSR2+ to be backported to even older GCN, like RT before it… If AMD were upfront and explained that just like XESS, the version of FSR4 that runs on non-supported hardware was in lower precision and is provided as-is, there'd be less backlash and there would be fewer accusations of half-baked features. Not letting older hardware get access to better quality (at the expense of performance) despite having put in the work to make it work… was bad. Letting everyone KNOW it works by accident was even worse. Just a bad look all around. And the lesson they'll likely learn from this is "don't get caught/let stuff leak". 😮💨 Reply
mitch074 qwertymac93 said: If AMD were upfront and explained that just like XESS, the version of FSR4 that runs on non-supported hardware was in lower precision and is provided as-is, there'd be less backlash and there would be fewer accusations of half-baked features. Not letting older hardware get access to better quality (at the expense of performance) despite having put in the work to make it work… was bad. Letting everyone KNOW it works by accident was even worse. Just a bad look all around. And the lesson they'll likely learn from this is "don't get caught/let stuff leak". 😮💨 Well, they did announce that they were looking into backporting those features – the accidental code drop was confirmation, because it's become an industry meme that AMD over-announce and under-deliver. After that, everybody got in a huff that AMD were splitting off RDNA1-2 on their own code path AKA "dropping support". That, when there was actual code evidencing how different RDNA 1-2 and RDNA 3-4 code paths were getting. So no – damned if they do, damned if they don't, I'll merely enjoy the eventual backporting of features to older hardware I'm still running on my backup systems, and give kudos where they're due, that the AMD driver team is actually trying to backport stuff where companies like Nvidia could give a rat's behind towards their customers' support. As for Intel I'd give them the benefit of the doubt if they weren't on the verge of bankruptcy and hadn't fired two-thirds of their Linux driver dev team. Edit : missing negative. Reply
Key considerations
- Investor positioning can change fast
- Volatility remains possible near catalysts
- Macro rates and liquidity can dominate flows
Reference reading
- https://www.tomshardware.com/pc-components/gpu-drivers/SPONSORED_LINK_URL
- https://www.tomshardware.com/pc-components/gpu-drivers/you-can-force-fsr-4-redstone-to-work-on-rdna-3-gpus-with-new-workaround-for-linux-systems-solution-requires-proton-compatibility-to-work-properly-gpu-drivers#main
- https://www.tomshardware.com
- AMD's legacy Ryzen 7 5800X3D chips now sell for up to $800, more than a new 9800X3D — AM4 chip costs twice as much as MSRP, as enthusiasts flock to old DDR4 mem
- French authorities arrest man for installing malware on a passenger ferry on behalf of 'a foreign power' — could have enabled external control of systems, inclu
- Grab this £138 2TB Acer Predator SSD before flash prices soar further — fast Gen 4 speeds make this an essential upgrade for your gaming PC
- China's new nation-spanning network eclipses 100 Gbps transferring 72 terabytes across 1,000km — experimental research network designed to connect thousands of
- Three year test shows OLED is significantly more reliable than LCD, with most lasting more than 10,000 hours — 20 TVs failed out of 102 tested
Informational only. No financial advice. Do your own research.