NVIDIA rumored to be preparing GeForce RTX 4080/4070 SUPER cards

From TPUs review:

Averaged over the 25 games in our test suite, at 4K, we find the card only 1.5% faster than 4080 non-Super Founders Edition, which is MUCH less than expected. While NVIDIA never said "+5%," I definitely expected more. It's not a power limit issue, running at max power yields another +1% only. These numbers are pretty constant across resolutions, and even with RT enabled, too. Looking at the individual games, the differences are 1 or 2 FPS, nothing you'd ever notice subjectively.

Looks like that Chinese leak was on the money.
 
It is a bit surprising how little +5.2% enabled core seem to do, maybe that the max this arch can do under 750 GB/s, Nvidia having calibrated that die size perfectly on the nose at launch.
 
Last edited:
It is a bit surprising how little +5.2% enable core seem to do, maybe that the max this arch can do under 750 GB/s, Nvidia having calibrated that die size perfectly on the nose at launch.
Based on the 4070ti Super, it might be the L2 cache makes a larger difference than initially expected.

Edit - to expand on this idea:

4070ti super -> 4080 super: 25% increase in L2 cache, 18% more performance.
4080 Super -> 4090: 12% increase in L2 cache, 13%more performance

Probably need to fill in cores and bandwidth into that comparison, but boy that L2 cache might be really really important to the card
 
Last edited:
Based on the 4070ti Super, it might be the L2 cache makes a larger difference than initially expected.
The rumors of even bigger cache on the 5090 despite not gaining in die size and cache not gaining much from better node could point to that (the large importance of the cache), specially if they achieve to go GDDR7 and still upgrade the cache a lot despite the massive bandwidth gain.

Maybe the ability to hit it with software that predict what to put in it or not make it pay more than before or something, RDNA 2 was a good success story in that regard as well.
 
Last edited:
The rumors of even bigger cache on the 5090 despite not gaining in die size and cache not gaining much from better node could point to that, specially if they achieve to go GDDR7 and still upgrade the cache a lot.

Maybe the ability to hit it with software that predict what to put in it or not make it pay more than before or something, RDNA 2 was a good success story in that regard as well.
Cache speed doesn’t scale up past TSMC 6N by any significant amount, you can get better density so more of it but it costs proportionately more so 128mb on 6N can cost less than half as much as 128mb on 3N for maybe a 3% performance improvement.

AMD, Apple, and Nvidia have been more than a little upset about that, and it’s one of the huge proposed selling features of Intels “L4” Adamantine cache as while it is slower it is significantly cheaper and can make up for that speed difference with sheer quantity.
 
4080 super goes out of stock at MSRP while 7900xt spotted at discount in mircocenter

https://www.hardwaretimes.com/nvidia-rtx-4080-super-price/

We might be heading towards a world where $1300 4080 super co-exists with a $650 7900xt

(Incidentally top RDNA 4 card, that could get released as soon as 6 months time, would be likely slightly slower than the 7900xt reminiscent of 5700xt vs Radeon VII scenario)
 
Last edited:
4080 super goes out of stock at MSRP while 7900xt spotted at discount in mircocenter

https://www.hardwaretimes.com/nvidia-rtx-4080-super-price/

We might be heading towards a world where $1300 4080 super co-exists with a $650 7900xt

(Incidentally top RDNA 4 card, that could get released as soon as 6 months time, would be likely slightly slower than the 7900xt reminiscent of 5700xt vs Radeon VII scenario)
People are probably picking these up as cheaper alternatives to 4090's for AI purposes. If the 7900 XT was good at AI, then people would grab them too. This whole thing reminds me of crypto.
 
Last edited:
Dell and others gaming PC builder:
https://www.newegg.com/p/pl?N=100897483 601432215&PageSize=96&page=1

Seem to have their hands on them.

Could be in the past (many things changed 2-3 time since) the high priced 4080 was a really good deal for inference or training, maybe that with a price cut it became the best deal "easy" to find :

GPUs_Ada_performance_per_dollar6.png
 
Back
Top