Saturday, February 22nd 2025

NVIDIA GeForce RTX 5090 Spotted with Missing ROPs, NVIDIA Confirms the Issue, Multiple Vendors Affected, RTX 5070 Ti, Too

TechPowerUp has discovered that there are NVIDIA GeForce RTX 5090 graphics cards in retail circulation that come with too few render units, which lowers performance. Zotac's GeForce RTX 5090 Solid comes with fewer ROPs than it should—168 are enabled, instead of the 176 that are part of the RTX 5090 specifications. This loss of 8 ROPs has a small, but noticeable impact on performance. During recent testing, we noticed our Zotac RTX 5090 Solid sample underperformed slightly, falling behind even the NVIDIA RTX 5090 Founders Edition card. At the time we didn't pay attention to the ROP count that TechPowerUp GPU-Z was reporting, and instead spent time looking for other reasons, like clocks, power, cooling, etc.

Two days ago, one of our readers who goes by "Wuxi Gamer," posted this thread on the TechPowerUp Forums, reporting that his retail Zotac RTX 5090 Solid was showing fewer ROPs in GPU-Z than the RTX 5090 should have. The user tried everything from driver to software re-installs, to switching between the two video BIOSes the card comes with, all to no avail. What a coincidence that we had this card in our labs already, so we then dug out our sample. Lo and behold—our sample is missing ROPs, too! GPU-Z is able to read and report these units counts, in this case through NVIDIA's NVAPI driver interface. The 8 missing ROPs constitute a 4.54% loss in the GPU's raster hardware capability, and to illustrate what this means for performance, we've run a couple of tests.

In the first test, "Elden Ring" at 4K UHD with maxed out settings and native resolution (no DLSS), you can see how the Zotac RTX 5090 Solid falls behind every other RTX 5090 we tested, including the NVIDIA Founders Edition, a de facto reference-design that establishes a performance baseline for the RTX 5090. The Zotac card is 5.6% slower than the FE, and 8.4% slower than the ASUS ROG Astral RTX 5090 OC, the fastest custom design card for this test. Officially, the Solid is clocked at 2407 MHz rated boost frequency, which matches the Founders Edition clocks—it shouldn't be significantly slower in real-life. The interesting thing is that the loss of performance is not visible when monitoring the clock frequencies, because they are as high as expected—there's just fewer units available to take care of the rendering workload.

A ROP (Raster Operations Pipeline) unit in the GPU processes pixel data, handling tasks like blending, antialiasing, render-to-texture, and writing final pixel values to the frame buffer. In contrast, a shading unit, aka "GPU core" is responsible for computing the color, lighting, and material properties of pixels or vertices during the rendering process, without directly interacting with the frame buffer, so the performance hit of the eight missing ROPs depends on how ROP-intensive a game is.
For example, in Starfield, the performance loss is much smaller, and in DOOM Eternal with ray tracing, the card actually ends up close to its expected performance levels.

We've also put the card through a quick 3DMark Time Spy Extreme graphics score run.
  • NVIDIA Founders Edition: 25439
  • Zotac Solid: 22621
  • Gigabyte Gaming OC: 26220
This should be a number that you can test easily for yourself, if you're one of the lucky RTX 5090 owners. The quickest way is definitely to just fire up GPU-Z and look at the ROP count number, it should be "176."

So far, we know only of Zotac 5090 Solid cards that are affected, none of our review samples from ASUS, Gigabyte, MSI, Palit, and NVIDIA exhibit this issue, all 5090 owners should definitely check their cards and report back.

This is an issue with quality assurance at both NVIDIA and Zotac. NVIDIA's add-in card partners (AICs) do not have the ability to configure ROP counts, either physically on the silicon, or in the video BIOS, and yet the GPU, its video BIOS, and the final product, cleared QA testing at both NVIDIA and Zotac.

We are working with Zotac to return the affected card, so they can forward it to NVIDIA for investigation. At this time Zotac was unable to provide a statement, citing the fluidity of the situation. As for possible fixes. We hope the issue is localized to a bug with the driver or the video BIOS, so NVIDIA could release a user-friendly BIOS update tool that can run from within Windows and update the BIOS of the affected cards. If, however, the ROPs were disabled at the hardware-level, then there's little that end-users or even AIC partners can do, except initiating a limited product recall for replacements or refunds. If the ROPs really are disabled through fuses, it seems unlikely that NVIDIA has a way to re-enable those units in the field, because that would potentially provide details to how such units can be reactivated on other cards and SKUs from the company.

Update 14:22 UTC:
Apparently the issue isn't specific to Zotac, HXL posted a screenshot of an MSI RTX 5090D, the China-specific variant of the RTX 5090 with nerfed compute performance, but which is supposed to have 176 ROPs. Much like the Zotac RTX 5090 Solid, it has 8 missing ROPs.

Update 16:38 UTC:
Another card has been found, this time from Manli.

Update 17:30 UTC:
ComputerBase reports that their Zotac RTX 5090 Solid sample is not affected and shows the correct ROP count of 176. This confirms that the issue isn't affecting all cards of this SKU and probably not even all cards in a batch/production run.

Update 17:36 UTC:
Just to clarify, because it has been asked a couple of times. When no driver is installed, GPU-Z will use an internal database as fallback, to show a hardcoded ROP count of 176, instead of "Unknown." This is a reasonable approximation, because all previous cards had a fixed, immutable ROP count. As soon as the driver is installed, GPU-Z will report the "live" ROP counts active on the GPU—this data is read via the NVIDIA drivers.

Update 19:18 UTC:
A card from Gigabyte is affected, too.

Update Feb 22nd, 6:00 UTC:
Palit, Inno3D and MSI found to be affected as well

Update Feb 22nd, 6:30 UTC:
NVIDIA's global PR director Ben Berraondo confirmed this issue. He told The Verge:
NVIDIAWe have identified a rare issue affecting less than 0.5% (half a percent) of GeForce RTX 5090 / 5090D and 5070 Ti GPUs which have one fewer ROP than specified. The average graphical performance impact is 4%, with no impact on AI and Compute workloads. Affected consumers can contact the board manufacturer for a replacement. The production anomaly has been corrected.
Very interesting—NVIDIA confirms that RTX 5070 Ti is affected, too.

While NVIDIA talks about "one ROP unit," this really means "8 ROPs" in our context. Many years ago, marketing decided that higher numbers = better, so they started to report the number of pixels that can be processed per unit, instead of the actual unit counts. So in this case, one hardware unit is disabled, which mean eight pixels per clock less can be processed, resulting in a loss of "8 ROPs".
Add your own comment

325 Comments on NVIDIA GeForce RTX 5090 Spotted with Missing ROPs, NVIDIA Confirms the Issue, Multiple Vendors Affected, RTX 5070 Ti, Too

#51
mxthunder
remekraWell, damn:





Seems like I'm missing a few as well. Are we sure it's not a bug? I mean my scores in 3dmark seem to be in line, now it's OCed so its even better. Maybe the spec in the database is wrong?


Oh well that would explain it.
Read above in thread, apparently its an error on TPU site, not bug with GPUz. 112 is the correct number.
Posted on Reply
#52
Vya Domus
bubbleawsomeThe 5090 already uses a slightly disabled version of GB202, so I’m having a hard time imagining defect rates are so high to kill another 8 ROPs.
You need to understand that the larger the chip is the yields go down dramatically, it's why AMD is trying to use chiplets for everything, any section that you can fuse off is an improvement.
Posted on Reply
#53
W1zzard
ImTheHollowManI'm just checking on my 5080 and apparently I'm missing almost 13% ROPs...
This was a mistake in GPU DB, the correct number is 112. I've updated our database
Posted on Reply
#54
Vayra86
bubbleawsomeThe 5090 already uses a slightly disabled version of GB202, so I’m having a hard time imagining defect rates are so high to kill another 8 ROPs. And even if they are, the 5080 uses a fully enabled GB203 so if they do want to launch a 5080TI Super or something eventually they’ll have to bin the 5090 anyway. They could start saving those dies now.
We're still looking at a 750mm2 die at very high density and >90M possible failure points. Anything above 550-600mm2 is really quite prone to yield effects.

Transistors 92,200 million Density 122.9M / mm² Die Size 750 mm²

Now... we've seen AMD chase the chiplet route... and they're now NOT producing a very large die RDNA4 card. Do the math. I think we might see those chiplets come back in the near future. Its the ONLY way to avoid these problems, when you've concluded the only way to move up is by scaling things up.
Posted on Reply
#55
W1zzard
zcyandrewI tried to submit my vbios to online database, but it says BIOS reading not supported on this device.
I just released a new build of GPU-Z that supports BIOS reading on Blackwell
Posted on Reply
#58
LittleBro
And people thought that Radeon RX 9070 launch was a disaster. Including me. I was wrong. Terribly wrong.

I was wondering what do they do with all those defective RTX 4090 and 5090 chips that are actually functional but only partially.
Since it's insanely huge monolithic die, defect rate is muc higher than with chiplet approach. Yield must be way worse than with RTX 5080 dies.

Before this I read that RTX 5070 and 5060 are delayed because of performance issues:
videocardz.com/newz/nvidia-geforce-rtx-5070-and-rtx-5060-mass-production-faces-delays

No other AIB cards reported to be impacted by the same issue yet? I'd expect PNY and Manli to be in Zotac's club as well.
Posted on Reply
#59
bubbleawsome
Vya DomusYou need to understand that the larger the chip is the yields go down dramatically, it's why AMD is trying to use chiplets for everything, any section that you can fuse off is an improvement.
Vayra86We're still looking at a 750mm2 die at very high density and >90M possible failure points. Anything above 550-600mm2 is really quite prone to yield effects.

Transistors 92,200 million Density 122.9M / mm² Die Size 750 mm²

Now... we've seen AMD chase the chiplet route... and they're now NOT producing a very large die RDNA4 card. Do the math. I think we might see those chiplets come back in the near future. It’s the ONLY way to avoid these problems, when you've concluded the only way to move up is by scaling things up.
I’m aware of area scaling and the fact that this gpu is quite a bit larger than even the 4090, but TSMC 4N is a pretty mature process. Like I said they could’ve turned these into a 5080Ti later, or even some slightly different professional grade GPU that consumers don’t usually see. For NVIDIA to be willing to scrape the bottom of the barrel and impact the reputation of their halo product for just 8 more ROPs feels like either yields are worse than they expected (otherwise why not just release vanilla 5090 with 168 from the start) or something else is going on. Maybe they did almost release the 5090 with 168 ROPs and some snuck into the release channel, but that seems like a major screwup too.
Basically what I’m saying is that anything that could make them need to disable 8 ROPs should’ve been known about well before the 5090 came out anyway. Super weird situation here.
Posted on Reply
#60
alwayssts
I sometimes wonder if anyone's ever put a AD103 and a GB203 under a microscope yet. :P

You know...Just to make sure?

I was honestly thinking about that the other day...so when you see something like this it kinda verifies something like that *could* happen. It never made sense AD103 has 2/3 of a shader cluster.
Posted on Reply
#61
Vya Domus
bubbleawsomeotherwise why not just release vanilla 5090 with 168 from the start
Who knows, maybe board partners are pissed because the supply is so low and they wasted all that money on marketing and promotion and so Nvidia came up with a solution to slightly boost the supply of chips.
Posted on Reply
#62
alwayssts
I've been waiting for somebody to say it, and they haven't, so I will bc well...gotta have some fun w/ this stuff:

When nVIDIA does it, it's a bug. When AMD does it, it's a feature!
Posted on Reply
#63
bubbleawsome
Vya DomusWho knows, maybe board partners are pissed because the supply is so low and they wasted all that money on marketing and promotion and so Nvidia came up with a solution to slightly boost the supply of chips.
Hm, wasn’t there recently a rumor that due to lower than expected b200 demand some wafers were going to be pushed to GB202 production instead? Wouldn’t it be funny if actual rush in 5090 supply came from simply using broken chips and lying about it.
Posted on Reply
#65
InVasMani
Pepperidge Farm Remembers when you would unlocked more ROPs on NVIDIA GPU's...

Posted on Reply
#66
Wuxi-gamer
It's such a relief to know that I'm not the only one who has this problem. Thank you very much for checking it out. Now waiting for a response from Zotac/Nvidia.
Posted on Reply
#67
Cheeseball
Not a Potato
Damn. More ZOTAC drama again. LOL

First was the shitty customer support throughout 2014 to 2020, then blatant price gouging and mining-promotions during COVID. Now its cut-down GPUs which may not directly be ZOTAC's fault since they don't fab them themselves, but they are involved since its only been seen on their boards.
Posted on Reply
#68
Darmok N Jalad
LOL, on eBay, these Zotac 5090s are listed for over $4000.
Posted on Reply
#69
R0H1T
Just 5k short of a meme then :D

Posted on Reply
#70
dyonoctis
Vayra86Calculated risk. These chips are expensive. What's 8 rops, hopefully customer won't notice, profit.

Its the same lame low-effort production method as what we've seen on 12VHPWR. Bare minimum, and actually, not even that. On halo products. You can't even make it up. What used to be hidden behind just the unified product lines, is now showing its true nature because these cards have crossed every line of common sense. High wattages to cover up zero progression; a new gen that offers literally nothing new, etc. A 5090 is nothing other than a cheap ass x60 except with more cores and chips around it. Everything Geforce is now in the low-effort budget segment for Nvidia.

Enjoy that thought while you pay >2000 for your graphics card.
Back in the days, it would have been called a RTX 5090 SE. Or RTX 5090 168.
Posted on Reply
#71
wNotyarD
dyonoctisBack in the days, it would have been called a RTX 5090 SE. Or RTX 5090 168.
In times of GTX, it would be a GTS. So maybe RTS 5090?
Posted on Reply
#72
HD64G
5090D for China, remained and sold as a full 5090. What a mess this gen has become with burnt connectors, fake MSRPs and untra-low stock and now this... :kookoo:
Posted on Reply
#73
AusWolf
I'll call it the 5090 GTX 970 Edition. The way you're meant to be played.
Posted on Reply
#74
Chomiq
Darmok N JaladLOL, on eBay, these Zotac 5090s are listed for over $4000.
Collector's item by now.
Posted on Reply
#75
bubbleawsome
HD64G5090D for China, remained and sold as a full 5090. What a mess this gen has become with burnt connectors, fake MSRPs and untra-low stock and now thos... :kookoo:
5090D ROP count is the same as the 5090
Posted on Reply
Add your own comment
Feb 22nd, 2025 09:33 EST change timezone

New Forum Posts

Popular Reviews

Controversial News Posts