Nvidia GeForce RTX 4090 Reviews

Hmm, bandwidth concerns in some scenarios already then if memory needs an overclock. Starting to doubt those rumors of a "4090ti", or at least doubt they'd end up with any super appreciable performance increase. Of course AMD successfully delivered overclocked memory in the severely limited edition 6900LC. I suppose if Nvidia wanted it enough they could do their own 600 watt tdp liquid cooled FE only version.
4090Ti has full cache (96MB instead of 72MB on 4090) 2k more shaders (for 18176), official boost clock of 2.75Ghz (2.95 typ gaming) to reach this sweet 100TFOPS marketing achievement, and faster 24Gb/s memory. Out of the box, it's 10~20% faster than 4090. Launch can be anytime. It will depend on Navi31 perf...
 
Effective clock is the same with undervolting for me.
underclockm6fws.png
undervoltopd4z.png
 
Last edited:
Undervolting is definitely interesting with this architecture. Testing various setups with Speed Way

Stock - 94 fps - 413W
OC (+230 core +1400 mem) - 101 fps - 428W
OC (same as above, undervolt to 1000mv) - 99 fps - 398W
OC (undervolt to 950mv) - 97 fps - 353W
OC (undervolt to 950mv but memory at +0) - 92 fps - 338W

So by undervolting and overclocking core/memory you can end up with better perf than stock with 60W less. But running at the same clock but lower voltage gives slightly lower results, like other people have already seen.
What about simply overclocking the memory without any change to voltages or core clock?
 
Undervolting is definitely interesting with this architecture. Testing various setups with Speed Way

Stock - 94 fps - 413W
OC (+230 core +1400 mem) - 101 fps - 428W
OC (same as above, undervolt to 1000mv) - 99 fps - 398W
OC (undervolt to 950mv) - 97 fps - 353W
OC (undervolt to 950mv but memory at +0) - 92 fps - 338W

So by undervolting and overclocking core/memory you can end up with better perf than stock with 60W less. But running at the same clock but lower voltage gives slightly lower results, like other people have already seen.
i dont know if anyone has mentioned it , but i assume there will be some clock stretching going on as you lower voltage, just like Zen 1/2/3/4 will do. So you have to be carefull and always measure performance :)
 
So they could have sold this card as a 350W part with a regular cooler instead of this behemoth slapped with coolers made for 600W?
 
So they could have sold this card as a 350W part with a regular cooler instead of this behemoth slapped with coolers made for 600W?
Of course they could. There was never remotely any need for such high power draw. Even at 270w, this would have been a fantastic performer.

We didn't even need to see the 4090 in action to know this would be the case.
 
Clearly the one thing that hasn't changed this gen is bandwidth... At least in Speed Way most of the gains in perf come from overclocking the memory. The cache probably helps mitigate the hit quite a bit.

I could see a 4090 Ti in 9-12 months with the full fat AD102 with higher binned/clocked memory, with 10-15% uplift from stock 4090.
Memory seems pretty easily overclockable on these GPUs. My memory is at 24Gbps... so a 4090Ti with 24Gbps memory seems very likely.
 
Of course they could. There was never remotely any need for such high power draw. Even at 270w, this would have been a fantastic performer.

We didn't even need to see the 4090 in action to know this would be the case.
Yea they definitely could have. The cooling on this thing is WAY overengineered for what's required... BUT I am loving that fact that my GPU never leaves the 50-60'C range, and is insanely quiet.

They could have made them a bit smaller to fit in more cases for sure.. but I'm honestly happy with the one I got.
 

Stock -> Memory OC Only (+1500) -> Core OC Only (+260) -> Both

Much bigger gains with memory OC.
I can't tell what is shown in that page - all four results appear to say the memory clock is the same (2235MHz - is that right?). And no frame rate or power - so hard to compare with what you showed before...

But it's interesting that you're apparently reporting that performance appears to scale best with memory clock in what seemed like a compute-limited benchmark.
 
Back
Top