AMD Radeon RX 9070 XT Reportedly Outperforms RTX 5080 Through Undervolting

[AleksandarK]

AMD's Radeon RX 9070 XT is demonstrating unexpected performance gains through aggressive undervolting, with overclocking specialists documenting significant improvements that push the GPU past NVIDIA's pricier GeForce RTX 5080 in specific benchmarks. Recent tests by Der8auer using a PowerColor Red Devil RX 9070 XT revealed a 10% frame rate increase in Cyberpunk 2077 at 4K Ultra settings by applying a -170 mV voltage offset while increasing the power target to 110%. This modification enabled the GPU to reach clock speeds of 3.36 GHz, compared to 2.90 GHz at stock settings, resulting in 66 FPS versus the RTX 5080's 65 FPS in identical testing environments. The undervolting phenomenon appears consistent across the product line, with YouTuber Alva Jonathan achieving similar 10% performance improvements on the standard RX 9070 using ASRock's Steel Legend model.

Both testers discovered that traditional core clock overclocking yielded negligible results, suggesting these factory-overclocked cards are already operating near their architectural limits. The voltage-frequency curve adjustments effectively lower the voltage required for higher frequencies. Memory overclocking proved counterproductive, with error correction mechanisms actually reducing in-game performance when pushed beyond stable parameters. These results come with important caveats—both tested units are premium variants with enhanced power delivery and cooling solutions that sell significantly above AMD's MSRP. The PowerColor Red Devil commanded a $200 premium over the RX 9070 XT's $599 launch price, while the ASRock Steel Legend carried a $90 markup over the RX 9070's base $550 MSRP. Even with these premiums, however, the high-end RX 9070 XT models remain approximately $200 less expensive than NVIDIA's RTX 5080 while delivering comparable rasterization performance after optimization, despite NVIDIA's ongoing advantages in ray tracing capabilities and software ecosystem.



View at TechPowerUp Main Site | Source

 

[Event Horizon]

Not surprised at all. Both 9070 and 9070 XT have a lot of room for finetuning.

 

[GhostRyder]

Well its always fun to see what can be done with a little tuning. Always funny in this day and age that lowering the voltage results in higher speeds lol.

 

[Naito]

It would be awesome to see TPU do their own tests regarding this info!

 

[LabRat 891]

Not surprised at all. Both 9070 and 9070 XT have a lot of room for finetuning.

Ever notice how all the modern chips clock considerably higher the cooler they can be kept?

I wonder if a Watercooled TEC, Chilled Water, or an Active Heat Pump could allow the 'meager' Navi 48 to clock and compete well outside its weightclass?


NGL, a part of me wants to see the RTX memes made real, but in Navi 48.

 

[cfenton]

Isn't this very silicon lottery dependent? It's good to see results like this, but we don't know if they are typical without a lot more data.

 

[londiste]

The 304W 9070XT which as this particular model is with an increased 330W power limit and increased further to 110%, resulting in a 360W card. A pretty heavily overclocked card gets 1fps over stock 5080 (running under its power limit, at 330W) in a test that seems to architecturally prefer RDNA4. Cool but not really that impressive. As der8auer says, you can also overclock 5080 by about the same 10%.

 

[docnorth]

It would be awesome to see TPU do their own tests regarding this info!

Yeah, but there is a caveat. TPU will try to optimise both AMD and Nvidia (as it should), I”m not sure everyone expects this.

 

[GodisanAtheist]

Isn't this very silicon lottery dependent? It's good to see results like this, but we don't know if they are typical without a lot more data.

-100%

Just like overclocking, undervolting is far from a guarantee. These chips are tuned to maximize usable dies at a certain point on the v/f curve. There are undoubtedly going to be a lot of chips out there that start having issues with even minor reductions in voltage. Ofc better binned chips are going to be able to operate at a higher clock with less voltage, but not all of them.

Another thing to keep in mind is that the chip might be stable in games x/y/z and then completely crash and burn in games a/b/c. Stability isn't a constant and different workloads will stress different components of a chip and they might not all always be stable with an undervolt.

 

[LabRat 891]

Isn't this very silicon lottery dependent? It's good to see results like this, but we don't know if they are typical without a lot more data.

At least going by Vega, Navi II, and Navi III (I never had a Navi I card), there's *way more* at play than meets the eye.

For example: I can 'push' my 7900 XTX to clock to 3.1Ghz+, fairly easy. It may even play a given game or stress test for hours without issue. BUT, one alt-tab out, a new overlay onscreen, or just closing the game, and the card crashes (as its trying to rapidly change clock/voltage states).

There's good reason AMD's cards are 'overvolted' out of the box.
It's the only way *currently* to guarantee a card is stable across all the dynamic points in the Power/Heat/Clock curve(s). Consequently, it *appears* that there's a lot of headroom when there really isn't.

 

[Scattergrunt]

There's good reason AMD's cards are 'overvolted' out of the box.
It's the only way *currently* to guarantee a card is stable across all the dynamic points in the Power/Heat/Clock curve(s). Consequently, it *appears* that there's a lot of headroom when there really isn't.

There is arguably still more room in them than some NVIDIA cards (blackwell / ada seemed to provide better OC / UV headroom tbf, esp blackwell) but mostly agreed yea

 

[remekra]

That's simply how OC works on AMD cards since I don't know when, at least RDNA1.

UV and increase power limit, clocks go up.

Problem is that in case of Radeons it's far harder to determine stability.
And there are heaps of people claiming at least for 7900XTX that they run at 1050mv or lower no issues, because they have run 3dmark for an 1 hour and played two games.
I could run 3d mark stress tests for hours on end on my 7900XTX when I had it and it would be stable then you open up Witcher 3 cast Aard by the water couple of times and boom crash.
Very random and unpredictable. Sometimes I wonder if people that are complaining that their system is not stable with Radeons are simply pushing them too far.
Because they get a notion that getting big OC is normal on them and they "stress" tested them so nah it couldnt be it.

In case of my 5080 once I pushed the clock 15MHz too high crashes would happen all the time and it's far easier to determine what is the limit.

 

[wolf]

Tuned vs Tuned would be more interesting and fair no? Also seems silicon lottery dependent. But hey I'm sure some people got cards that clock up super well.

Just like some 5080's are tickling a 4090's undercarriage with tuning.

 

[sethmatrix7]

The 304W 9070XT which as this particular model is with an increased 330W power limit and increased further to 110%, resulting in a 360W card. A pretty heavily overclocked card gets 1fps over stock 5080 (running under its power limit, at 330W) in a test that seems to architecturally prefer RDNA4. Cool but not really that impressive. As der8auer says, you can also overclock 5080 by about the same 10%.

It matches the performance of a card often found for double its price. I’d say that’s pretty impressive.

 

[TheinsanegamerN]

All this just makes me wish even more that AMD made a 9080xt.

The 304W 9070XT which as this particular model is with an increased 330W power limit and increased further to 110%, resulting in a 360W card. A pretty heavily overclocked card gets 1fps over stock 5080 (running under its power limit, at 330W) in a test that seems to architecturally prefer RDNA4. Cool but not really that impressive. As der8auer says, you can also overclock 5080 by about the same 10%.

Yeah you can, but the fact a $600 card can be tuned to tickle a $1000 card is still impressive. It's the kind of thing we used to get excited over in the fermi era.

 

[Darmok N Jalad]

That's simply how OC works on AMD cards since I don't know when, at least RDNA1.

UV and increase power limit, clocks go up.

Problem is that in case of Radeons it's far harder to determine stability.
And there are heaps of people claiming at least for 7900XTX that they run at 1050mv or lower no issues, because they have run 3dmark for an 1 hour and played two games.
I could run 3d mark stress tests for hours on end on my 7900XTX when I had it and it would be stable then you open up Witcher 3 cast Aard by the water couple of times and boom crash.
Very random and unpredictable. Sometimes I wonder if people that are complaining that their system is not stable with Radeons are simply pushing them too far.
Because they get a notion that getting big OC is normal on them and they "stress" tested them so nah it couldnt be it.

In case of my 5080 once I pushed the clock 15MHz too high crashes would happen all the time and it's far easier to determine what is the limit.

Games are a different beast, for sure. It’s probably because the scenes are far more complex and have random spikes that might coincide with background loading, who knows.

It’s the same for my RX 7600. I can get it over 3GHz in Forbidden West, but Hogwarts Legacy is a crash fest.

 

[Fouquin]

That's simply how OC works on AMD cards since I don't know when, at least RDNA1.

Since Fiji in 2015. That's how AMD made the R9 Nano a reality; it's just a Fury X with an extremely strict power and voltage limit. Nothing more. If you drag the power limit back up and balance the voltage around 1v you get a card that is 1:1 with a Fury X but at like 225W instead of 300W. Vega was way overvolted by default which allowed the Vega 64 to boost decently high, but on Vega 56 it was entirely unnecessary; hence the plethora of 1.05-1.08v tuned cards running 1650MHz+ once the power limit is cranked to the sky. Arguably it was not needed for Vega 64 but a lot of the Vega samples struggled to reach their boost targets early on so I believe they left the 1.2v boost profiles in place to ensure stability across production.

 

[N/A]

that means 25% overvoltage by default. i have to give it 450W TGP and let it run wild.

 

[AusWolf]

That's embarrassing... For Nvidia.

 

[Epaminombas]

You have to be very brave to play with GPU voltage and clock to void the warranty.

Then the people who bought it will be crying because they saw that it is safe to do

 

[DaemonForce]

Isn't this very silicon lottery dependent? It's good to see results like this, but we don't know if they are typical without a lot more data.

How about vram overclocking? I wouldn't place any bets on SKHynix cooking up any screamers.
Infineon, Micron and Samsung have been my favorites, in that order.

All this just makes me wish even more that AMD made a 9080xt.

Yeah you can, but the fact a $600 card can be tuned to tickle a $1000 card is still impressive. It's the kind of thing we used to get excited over in the fermi era.

>UDNA time
>AMD just casually ships a 32GB GDDR7 screamer
>Calls it the RX 9080XT

>"But Lisa, why the weird name?"
>"Well we were just thinking of our competition and a fun way to send them to the funny farm. 90 class in raster, 80 class in RT..."

 

[tpa-pr]

I have passed the above onto my friend who is desperate for a Sapphire 9070 XT. He is currently salivating at the prospect of getting one.

It's fascinating how RDNA4, a refresh of the previous gen, turned out to be quite the banger. I guess part of it is going back to the tried-and-tested monolithic design? I do hope they continue to explore chiplets for UDNA though.

 

[N/A]

I have passed the above onto my friend who is desperate for a Sapphire 9070 XT. He is currently salivating at the prospect of getting one.

It's fascinating how RDNA4, a refresh of the previous gen, turned out to be quite the banger. I guess part of it is going back to the tried-and-tested monolithic design? I do hope they continue to explore chiplets for UDNA though.

Not chiplet is contributing to latency and infinity band, but i recon it's still there, just brings them closer together.
Mainly because 33% more ROPs than 7800 XT, and not very surprising 33% more performance.
Save your money for UDNA if I were him. due to the N3 shrink. that's 50% more shaders in the same area.
And more precisely, if the 128 CU theoretical flagship would have been 529mm2, it shrinks to 400mm2 just perfectly.

 

[wolf]

That's embarrassing... For Nvidia.

I'm sure Jensen will cry all the way to the bank, given he's selling a 5070 as a 5080 and can't make enough of them to satisfy demand.

 

[Bjorn_Of_Iceland]

Not surprised. Some of the benchmarks hinted it was reaching 5080 levels of performance back when the leaks were left and right.