Volt-modded RX 7900 XTX Hits 3.46 GHz, Trades Blows with RTX 4090

[btarunr]

An AMD Radeon RX 7900 XTX graphics card is capable of trading blows with NVIDIA GeForce RTX 4090, as overclocker jedi95 found out. With its power limits unlocked, the RX 7900 XTX was found reaching engine clocks as high as 3.46 GHz, significantly beyond the "architected for 3.00 GHz" claim AMD made in its product unveil last Fall. At these frequencies, the RX 7900 XTX is found to trade blows with the RTX 4090, a segment above its current segment rival, the RTX 4080.

Squeezing 3.46 GHz out of the RX 7900 XTX is no child's play, jedi95 used an Elmor EVC2SE module for volt-modding an ASUS TUF Gaming RX 7900 XTX, essentially removing its power-limit altogether. He then supplemented the card's power supply, so it could draw as much as 708 W (peak), to hold its nearly 1 GHz overclock. A surprising aspect of this feat is that an exotic cooling solution, such as liquid-nitrogen evaporator, wasn't used. A full-coverage water block and DIY liquid cooling did the job. The feat drops a major hint at how AMD could design the upcoming Radeon RX 7950 XTX despite having maxed out the "Navi 31" silicon with the RX 7900 XTX. The company could re-architect the power-supply design to significantly increase power limits, and possibly even get the GPU to boost to around the 3 GHz-mark.



View at TechPowerUp Main Site | Source

 

[nguyen]

only 17% gain in CP2077, meanwhile 4090 is on average 25% faster than 7900XTX.

So yeah, trade blows in CP2077 and Timespy only LOL

 

[Space Lynx]

only 17% gain in CP2077, meanwhile 4090 is on average 25% faster than 7900XTX.

So yeah, trade blows in CP2077 and Timespy only LOL

all benchmarks are nonsense anyway imo, most of us, maybe not you 4090 owners, but most of us don't mind playing games at a mixture of medium to high settings to increase fps, and if you own a x3d cache cpu, a lot of games get even more bonus fps from that, but only when turned down to a mix of a medium to high settings. so many sites and youtubers never show these things though, so i have learned to take any benchmarks as a grain of salt. i still enjoy reading w1zz's and watching gamersnexus though when new gpu's come out.

edit: thats not a slight to 4090 owners btw, if i had a better paying job id more than happy buy a 4090 and max out games. lol

 

[phanbuey]

Now let's see what a volt modded 4090 can do!

 

[zo0lykas]

Now let's see what a volt modded 4090 can do!

Empty your wallet lol

 

[Zubasa]

Now let's see what a volt modded 4090 can do!

Note that der8auer promptly kiiled his card afterwards.

 

[Claudio]

only 17% gain in CP2077, meanwhile 4090 is on average 25% faster than 7900XTX.

So yeah, trade blows in CP2077 and Timespy only LOL

the timespy is misleading. He used a 13900k with the 7900xtx which raised the overall score by a lot, since an AMD 7950 was used with the 4090. So if you look at the Grapichs score, the 4090 is always ahead

 

[Dr. Dro]

I find it an interesting project, but if anything, it's testament to Nvidia's architectural superiority this time around. The RTX 4090 is simply uncontested, and in response to an eventual clock-buffed, power-hungry 7950 XTX, NVIDIA can simply enable the remaining partitions and cache slices in the AD102 to vaporize whatever untapped potential N31 may have had left.

 

[Hxx]

I find it an interesting project, but if anything, it's testament to Nvidia's architectural superiority this time around. The RTX 4090 is simply uncontested, and in response to an eventual clock-buffed, power-hungry 7950 XTX, NVIDIA can simply enable the remaining partitions and cache slices in the AD102 to vaporize whatever untapped potential N31 may have had left.

the 4090 is also $6-$900 more depending on which models you are comparing. To me is when the 7900xtx beats it in some games is how good value this card is. Otherwise i would expect a much more expensive card to be faster and more efficient across the board.

 

[Dr. Dro]

the 4090 is also $6-$900 more depending on which models you are comparing. To me is when the 7900xtx beats it in some games is how good value this card is. Otherwise i would expect a much more expensive card to be faster and more efficient across the board.

The 4090 is more expensive, yes, but remember: it holds the distinct performance crown, serves as Nvidia's halo, is far more power efficient, and it's not even stretched out to its full potential. In fact, it's the worst flagship NVIDIA has ever released in terms of die quality and relative amount of enabled units. There is simply no way that any Navi 31 design can perform at the AD102 level when presented with fully enabled silicon v. fully enabled silicon.

The only games where the 7900 XTX equals the 4090 are games where there are problems with CPU optimization or something seriously wrong with the Nvidia driver (yes, it can happen). In any game which performs normally and can leverage the hardware correctly, it doesn't stand a chance.

 

[nguyen]

7900XTX after running within an inch of its life

 

[JustBenching]

the 4090 is also $6-$900 more depending on which models you are comparing. To me is when the 7900xtx beats it in some games is how good value this card is. Otherwise i would expect a much more expensive card to be faster and more efficient across the board.

Uhm, that's silly. The 4070ti beats the 7900xtx in some games as well. Is that a testament to how great a value it its?

 

[evernessince]

I find it an interesting project, but if anything, it's testament to Nvidia's architectural superiority this time around. The RTX 4090 is simply uncontested, and in response to an eventual clock-buffed, power-hungry 7950 XTX, NVIDIA can simply enable the remaining partitions and cache slices in the AD102 to vaporize whatever untapped potential N31 may have had left.

This is not a good take IMO. This is akin to saying Zen 1 was a terrible architecture because it wasn't the most efficient or fastest product on the market. You are ignoring the fact that the 7900 XTX only has a graphic die size of around 300mm2. The 4090 on the other hand is a 608mm2 size die. As we all know, breaking a die down into smaller pieces drastically reduces cost and increases yields, particularly when we are talking about large dies like GPUs.

That's aside from the other advantages of chiplets, like being able to stack cache, modularization of parts of the GPU (which allowed AMD to increase the density of cores in it's latest GPUs), exceed the reticle limit, improved binning, and more.

Architecturally speaking Ada is Ampere on a much better node with extra cache taped on and not the fancy stacked cache either. RDNA3 simply being the first dGPU based on a chiplet design makes it more impressive than adding more cache to a monolithic die as Nvidia has done. The only shame is that AMD didn't go for the throat with pricing like they did with Zen 1. The 4090 is an impressive card but the architecture itself is more a beneficiary of node improvements than itself being anything revolutionary.

 

[Hxx]

Uhm, that's silly. The 4070ti beats the 7900xtx in some games as well. Is that a testament to how great a value it its?

Sure whatever game that is then it’s a good buy compared to the xtx - let’s say that’s the game or games you’re mostly playing would u pay less for more performance or more for less performance ?
no different than someone buying a card to mostly play COD where the xtx is faster than a 4090 by quite a bit

 

[Outback Bronze]

Still, good to see people overclocking the snot out of the less expensive card to keep up with the more expensive one.

Love to see a volt-modded 4080 take on a 4090.

Just like the good old days. Buying the cheaper one and getting the same performance as the more expensive ones. I feel those days are gone sadly.

 

[Vayra86]

I find it an interesting project, but if anything, it's testament to Nvidia's architectural superiority this time around. The RTX 4090 is simply uncontested, and in response to an eventual clock-buffed, power-hungry 7950 XTX, NVIDIA can simply enable the remaining partitions and cache slices in the AD102 to vaporize whatever untapped potential N31 may have had left.

Architectural superiority?!

Did you consider this?

vs

And let's not forget the 5nm GCD is only 300mm2 as well.

Ada is not superior, its more dense, a bigger die and a LOT more expensive to bake. Most other metrics that originate from the hardware are much the same, such as power consumption / W/frame. That 20% ish gap on raster translates almost directly in the transistor count gap as well. What else has Nvidia got? Featureset. Not hardware or architecture. That's why they push DLSS3 so hard.

This is the equivalent of early day Ryzen vs Intel monolithic, and its clear AMD is building competitive advantage if you look at die size / perf.

This is not a good take IMO. This is akin to saying Zen 1 was a terrible architecture because it wasn't the most efficient or fastest product on the market. You are ignoring the fact that the 7900 XTX only has a graphic die size of around 300mm2. The 4090 on the other hand is a 608mm2 size die. As we all know, breaking a die down into smaller pieces drastically reduces cost and increases yields, particularly when we are talking about large dies like GPUs.

That's aside from the other advantages of chiplets, like being able to stack cache, modularization of parts of the GPU (which allowed AMD to increase the density of cores in it's latest GPUs), exceed the reticle limit, improved binning, and more.

Architecturally speaking Ada is Ampere on a much better node with extra cache taped on and not the fancy stacked cache either. RDNA3 simply being the first dGPU based on a chiplet design makes it more impressive than adding more cache to a monolithic die as Nvidia has done. The only shame is that AMD didn't go for the throat with pricing like they did with Zen 1. The 4090 is an impressive card but the architecture itself is more a beneficiary of node improvements than itself being anything revolutionary.

Oh lol, missed this one

 

[JustBenching]

Architectural superiority?!

Did you consider this?

View attachment 296480 vs View attachment 296482

And let's not forget the 5nm GCD is only 300mm2 as well.

Ada is not superior, its more dense, a bigger die and a LOT more expensive to bake. Most other metrics that originate from the hardware are much the same, such as power consumption / W/frame. That 20% ish gap on raster translates almost directly in the transistor count gap as well. What else has Nvidia got? Featureset. Not hardware or architecture. That's why they push DLSS3 so hard.

This is the equivalent of early day Ryzen vs Intel monolithic, and its clear AMD is building competitive advantage if you look at die size / perf.


Oh lol, missed this one

A big part of the transistors in the 4090 is the huge cache, which nvidia claims specifically helps with rt workloads. And in those the performance difference between the 2 cards is massive, much bigger than the transistor count. Also you can check 4080 vs the 7900xt,the 4080 is much smaller and much faster

 

[Vayra86]

A big part of the transistors in the 4090 is the huge cache, which nvidia claims specifically helps with rt workloads. And in those the performance difference between the 2 cards is massive, much bigger than the transistor count. Also you can check 4080 vs the 7900xt,the 4080 is much smaller and much faster

The 7900XT has disabled units, so it doesn't actually have/use 57,700M transistors obviously.
The 4080 still has 45,900M
and oh look:
57,7 : 45,9 = 1,2570
Ergo 25% gap

The cache just makes sure the 4090 doesn't choke as hard on RT workloads as the rest of the stack, so it shows in a handful or heavily RT focused titles, yay
The rest is just cold hard numbers, no avoiding it.

 

[JustBenching]

The 7900XT has disabled units, so it doesn't actually have/use 57,700M transistors obviously.
The 4080 still has 45,900M
and oh look:
57,7 : 45,9 = 1,2570
Ergo 25% gap

View attachment 296483

But the xtx is bigger - has more transistors - and is much slower in rt? How is this not a clear indication of architectural superiority. I mean is obvious that if the while die of the 4080 or the 4090 was used for raster performance the gap would be massive in raster, no?

 

[Vayra86]

RT is simply not a valid performance metric at this point, but you do you. The better half of that perf advantage is Nvidia dropping bags of money left and right to get their RTX marketing story in the game.

 

[JustBenching]

RT is simply not a valid performance metric at this point, but you do you.

Why not? You are comparing architectures. Whether or not you think rt is useless is completely irrelevant here.
I would argue raster is not a valid performance metric...

 

[Vayra86]

Why not? You are comparing architectures. Whether or not you think rt is useless is completely irrelevant here.
I would argue raster is not a valid performance metric...

Good luck gaming without raster then. You've got two games ahead of you, both running at sub 30 FPS on a 4090.

That is why not. I'm gonna leave it here

 

[JustBenching]

Good luck gaming without raster then. You've got two games ahead of you, both running at sub 30 FPS on a 4090.

That is why not. I'm gonna leave it here

But it doesn't matter, you are comparing architectures, you can't leave rt out of it.

That's like comparing a 16 core cpu to an 8core cpu in single threaded performance and claiming the 2nd one is better architecturally cause it has a much smaller die. I mean come on....you know you are wrong.

 

[Vayra86]

But it doesn't matter, you are comparing architectures, you can't leave rt out of it.

That's like comparing a 16 core cpu to an 8core cpu in single threaded performance and claiming the 2nd one is better architecturally cause it has a much smaller die. I mean come on....you know you are wrong.

The problem with RT is that we have no sane metric for it. The performance drop you get - in both camps - depends heavily on the amount of RT used. So we can compare architectures, and sure, Nvidia HAS more perf on offer there, but its misty as fuck. Fact is, every single GPU chokes hard on RT, today, with perf loss north of 40% just for mild IQ improvements. Its not sustainable 'as is'.

The reason I say this is because we're still in the early adopting days. There is no transparency here wrt which architecture will remain best fit going forward. What we are seeing though is that engines deploy lighter / more adaptive technologies to cater to different performance levels. Nvidia is capable of leveraging their proprietary approach today, but that advantage is going to end soon - and its not an architectural advantage so much as an optimization exercise with developers.

When RT gets more mainstream and there is more consensus on how it gets used, we can get a much better view on what the architectural impact is on RT perf. Today it simply is bonus, at best, and many gamers will be turning it off, especially lower down the stack.

The analogy to core counts on CPUs though... lol. There is a lot wrong with that - if games would have been primarily single threaded, guess what, the 8 core CPU is definitely the more competitive product and if it has, say, a fast cache to accelerate gaming workloads and topped out at 8 cores, while beating high core count adversaries in the other camp on a smaller footprint, then yes, it is architecturally stronger, at least for that use case. So I guess if you buy into the RT fairy tale, you can defend Ada is stronger for your use case right now.

But there is the long term perspective here, and with architectures that one matters. Nvidia I'm sure has a plan going forward, but the fact is, 600mm is the end of the road for Geforce products, and they're on it, while giving that same SKU a very high power limit while on a very efficient node. All I see here is another Intel, to be fair, with the added point of using RT/DLSS to avoid the death of monolithic instead of a big little approach

 

[TheoneandonlyMrK]

only 17% gain in CP2077, meanwhile 4090 is on average 25% faster than 7900XTX.

So yeah, trade blows in CP2077 and Timespy only LOL

Let's see you Beat a 1Ghz OC IE the effin point of the OP not your precious.