If this is when AMD move to chiplets in GPUs, I could actually see larger than normal performance ceiling rise on the high-end- but I wouldn't believe for a moment that performance in the midrange or perf/$ in general will go up by anything remotely close to 3x. If the top-end RX 7000 or whatever is somehow 2.5-3x as fast as 6900XT by virtue of chiplet scaling on top of IPC & clock gainz, it'll also be an even higher performance tier (7990?) and cost more accordingly. Feel free to quote the entirety of this post next year to see how I did lol.
I'm thinking along those lines tooGoing by current msrp
6900 XT > $999
7990 XT > $2499
Last paper I read from NVidia from like 2 years ago was they were working on how to lower the latency between the IO die and the chiplets to improve frame time issues. I have to assume that they all had similar issues and have been working on it all this time. 3 years seems like a reasonable timeframe to tackle this issue.Eh, they have a different solution going. Basically the system only sees one GPU, which is the I/O die. The I/O die tells the chiplets what to do, and takes the info they handle and present it to the system. Essentially, it's just one big GPU even if it's hosted on multiple physical dies.
I don't know if they work like one big die or if they split up roles, like SFR back in the day. But however it'll work, it's a black box as far as the system is concerned.
geez I hope not... NV and AMD spent the entire 2010's insisting that "~this time~ AFR frame pacing, scaling, and compatibility is fixed!" and it never really was. I'm hoping for more like the arrangement on Ryzen 3000/5000 where the IO die holds the last-level cache and scheduler and treats the compute dies as a single pool of threads.Makes it sound like AFR is back on the table...
I'm hoping for more like the arrangement on Ryzen 3000/5000 where the IO die holds the last-level cache and scheduler and treats the compute dies as a single pool of threads.
And you cant buy one