Nvidia GeForce RTX 4090 Reviews

Microbenchmarks of RTX 4090 at Chips and Cheese: Microbenchmarking Nvidia’s RTX 4090. The tests suggest Nvidia's most significant changes with the Ada architecture can be found in the memory infrastructure to enable Ada to scale up beyond Ampere notwithstanding the lack of appreciable increases in VRAM bandwidth.

On another note, I picked up my 4090 FE today. It's built like a brick (in a good way) and I had to remove the side panel from my Lian Li O-11 Dynamic to accommodate the Nvidia adapter. I'm eagerly awaiting CableMod's 180 Degree adapter. I haven't had much time to play with the card other than to do a quick run of Cyberpunk at Max (Psycho) settings. It's very quiet though I did notice a bit of coil whine when the GPU clocks over 2700mhz or so. At a cursory glance, changing the power limit from 133% to 100% to 80% didn't affect performance very much at all (but did lower power by 100w from min to max). I'm currently running a 1440p 165hz monitor.... could be I'm hitting a CPU bottleneck on my 5900x and/or relatively slow system memory (DDR4 3200 16-18-18-36).
Does changing the power limit control the coil whine? Honestly curious, and congrats on the card. :)
 
It's very quiet though I did notice a bit of coil whine when the GPU clocks over 2700mhz or so. At a cursory glance, changing the power limit from 133% to 100% to 80% didn't affect performance very much at all (but did lower power by 100w from min to max). I'm currently running a 1440p 165hz monitor.... could be I'm hitting a CPU bottleneck on my 5900x and/or relatively slow system memory (DDR4 3200 16-18-18-36).
The little coil whine you hear might be attributable to the removal of the side panel. At some forums 4090 users with the same card mentioned hearing the whine while others did not. They are still trying to figure out if it is related to certain types of PSU/rating. It's also possible the coil whine is just like any previous gen card ... some have it and others don't.
 
Microbenchmarks of RTX 4090 at Chips and Cheese: Microbenchmarking Nvidia’s RTX 4090. The tests suggest Nvidia's most significant changes with the Ada architecture can be found in the memory infrastructure to enable Ada to scale up beyond Ampere notwithstanding the lack of appreciable increases in VRAM bandwidth.

On another note, I picked up my 4090 FE today. It's built like a brick (in a good way) and I had to remove the side panel from my Lian Li O-11 Dynamic to accommodate the Nvidia adapter. I'm eagerly awaiting CableMod's 180 Degree adapter. I haven't had much time to play with the card other than to do a quick run of Cyberpunk at Max (Psycho) settings. It's very quiet though I did notice a bit of coil whine when the GPU clocks over 2700mhz or so. At a cursory glance, changing the power limit from 133% to 100% to 80% didn't affect performance very much at all (but did lower power by 100w from min to max). I'm currently running a 1440p 165hz monitor.... could be I'm hitting a CPU bottleneck on my 5900x and/or relatively slow system memory (DDR4 3200 16-18-18-36).
congrats on the card, enjoy it, you have a GPU for a couple of generations. With that amazing hardware a good 4K monitor or TV would be perfect, specially now with reconstruction techniques such as DLSS and XeSS, you can have a very silent GPU resting on its laurels lol at 4K.

In regards to coil whine, afaik I've never experienced it on a computer of mine, although I could hear it in computers I opened. I heard that it might go over time as you give the GPU some use, but maybe you are in time to ask for a replacement or look for another model. Maybe it's not a model thing, but rather some cards behave differently.
 
I did several runs through Port Royal to find an approximate threshold for the whine I'm hearing by changing GPU power and core offsets in Afterburner.

Power Limit \ Core Clock Offset \ Avg. Clock \ Score \ Whine

133% -- +100 -- 2805mhz -- 412W -- 25803 - Yes
133% -- 0 -- 2715mhz -- 25425 - Yes
100% -- +100 -- 2805mhz -- 25671 - Yes
100% -- 0 -- 2715mhz -- 25312 - Yes
90% -- +100 -- 2790mhz -- 25574 - Yes
90% -- 0 -- 2700mhz -- 25287 - Yes
80% -- +100 -- 2677mhz -- 25131 - Yes
80% -- 0 -- 2598mhz -- 24842 - Yes
70% -- +100 -- 2553mhz -- 24274 -- Yes but much less... almost certainly not audible with side panel on.
70% -- +0 -- 2484mhz -- 23967 - Same

While I was running these tests, I stuck my head inside my PC with my ear (dangerously) close to the top fan on the GPU. One, the "pass through" fan on the 4090 is impressively quiet. Two, I became less convinced the whine was from the GPU. I ran the 133% power limit again, pulled off the back of my case, and Lo! The whine seems to be coming from my Corsair HXi 850 power supply. I've never heard a sound from it or known its fan to spin (even coming from an EVGA 3080 FTW), but the PSU fan was spinning slowly and the buzzing seemed to come from within (not from the PSU fan). So, it appears my PSU has whine that occurs when it is pushed harder than it was by the 3080. It probably isn't a coincidence that the whine mostly disappeared when the power draw of the 4090 was limited to 70% and was pulling about as much power as my 3080.

My sincerest apologies to Jensen and Nvidia's engineers for accusing the card of coil whine :)! Fortunately, a PSU is more easily replaced than a 4090.
 
I have a few more observations having tinkered with my 4090's power connector and power and core settings while doing Port Royal runs:

1) I unplugged one of the 4 PCIe power cables from the adapter (giving the card access to 450W max), which had no appreciable hit on performance in PR. Afterburner automatically detects and limits the power limit slider to 100% i.e. 450W. Because I never saw the 4090 go anywhere near 450W in any PR run (even when the power limit was set to 133% with all 4 PCIe cables in), it makes sense that there was no performance change.

2) I unplugged 2 of the PCIe cables from the adapter, giving the card 300W of power. The system booted just fine (I could hear Windows doing its thing) but I couldn't get my either of my two monitors to display a signal (a 1440p 165hz Asus as main and an LG 4k 60hz both hooked through displayport). Weird.

3) You won't be surprised that the last bit of performance costs dearly in power. At 100% power (so 450W available) and a +200 core offset (avg. clock 2881mhz, and it crashed at +250 offset), I scored 26079 and used around 410W on average. At 80% power and a +250 core offset, I hit 25532 using only 358W. At 75% and +250, hit 25222 at 335W. And, at 70% with +250, I got 24583 using 313W. That's what, a 6% performance loss for a 23% power savings? With the usual caveats of course: limited testing runs, limited to Port Royal, using my system, etc.

4) I hate my Corsair power supply now that it whines. It's a platinum rated 850W supply, and I'm not even hitting 700W when I have the 4090 at 100%. I suppose I'll need to wait for Seasonic (they're still the top shelf, eh?) to release its ATX 3.0 supplies with the PCIe 5.0 connector later this year or early next year. Sad!
 
I can't remember the last time we had a single GPU that was so CPU limited as the 4090 is, even on the latest CPU's and platforms.

There are some very specific games I can think of that have always been CPU limited on older GPU generations but nothing to the extent of or as common as the 4090 is.
 
Last edited:
I can't remember the last time we had a single GPU that was so CPU limited as the 4090 is, even on the latest CPU's and platforms.

There are some very specific games I can think of that have always been CPU limited on older GPU generations to nothing to the extent of or as common as the 4090 is.
PowerVR PCX1 and 2 needed a pretty strong CPU to give their best :eek:
 
Heh. I remember getting my 3Dfx Deltron Realvision Flash 3D Voodoo 1 with my Pentium 120Mhz (as I thought at the time that the VideoLogic Apocalypse 3D needed a 166Mhz + CPU to not be CPU limited). Low and behold.... Shadows of the Empire and Extreme Assault was very CPU limited with Voodoo 1 and 120 Mhz. Had to updgrade to P200Mhz MMX to eleveate the CPU bottleneck.

Now with my MSI 4090 Suprim X is significanlty bottlenecked by my 9900KS. Going to have to get a 13900K to allow high FPS gaming again.

I am currently having to use DLDSR to make it somewhat GPU limited again!
 
I've been fiddling around with F@H and got a good chuckle. I've folded on and off since the days of, I believe, my SLI'd GTX 470s and in that time earned a little over 46 million points, which ranks me around 46,000 out of 2.9 million. On the project I'm running now on the 4090, it says I'll earn around 26-27 million points per day. I knew inflation was insidiously sinking its teeth into everything around us, but even F@H!!! I don't know how accurate the point per day counter is, but I got a kick out it.

For what it's worth, F@H is another program that does not fully saturate the 4090 on my system. At 100% (450W) power limit and +150 core offset (core pegged at 2850 mhz), the Nvidia Performance Overlay indicates 95% GPU utilization (with ~ 7% CPU on my Ryzen 5900x) but only drawing ~353W. After over 30 minutes, the core temp hovers around 65-66c. I threw +800W on the memory to see what would happen. The power draw actually went down to ~348W, but the F@H project had changed by this time, so that may explain the difference in power. There doesn't seem to be an appreciable change to the PPD estimate.

At 60% power limit with the same +150 core offset (core bounces around far more often, now typically between 2445-2565mhz with spikes into the 2800mhz range), the power draw floats around 256W and the core temp around 56c. The point counter indicates between 23-25 million points per day (I expect the wider range arises from the bouncing core clock).

While running the 4090 at 100% and folding on GPU/CPU (5900x) simultaneously, my PSU is reporting between 630-670W "Power In" and 566-590W "Power Out". Of course, the PSU is definitely doing its coil whine right now, so PSU and I aren't on speaking terms.
 
Even though the product costs more than $1,500, many people have bought the NVIDIA GeForce RTX 4090. Note that 100,000 units do not equal the total number of GPUs made or sold. Instead, it is the number of chips that were sent to partners so they could make RTX 4090 models.
 
Back
Top