Review Nvidia GeForce RTX 5090 Founders Edition review: Blackwell commences its reign with a few stumbles

[helper800]

5090 is actually 4090 ti renamed . the 27% increase is just 30% of more transistor counts. no generational leap here.

Ada Lovelace and Blackwell architectures are distinctly different. The main similarity is that they are made using a similar process node.

 

[JarredWaltonGPU]

I'm most certainly waiting for DLSS review deep dive, on the base level it is clear that what you say is what is happening.

But what I wonder is image quality/artifacts comparison between Performance/Balanced and Quality modes, because maybe now using Balanced/Performance modes would result in better image quality than DLSS2 at Quality mode had - which should be a performance increase while keeping up the good reproduction quality.

I've been poking at this a bit in Cyberpunk 2077. I don't have comparable data right now for other GPUs, but DLSS Transformers definitely looks better. That will, as always, vary by the game and scene you're looking at, but let me see if I can drop some images to illustrate.

4K, DLSS Quality Transformers, FrameGen 2X (basically what we had with 40-series but using a new model):


4K, DLSS Quality Transformers, MFG 3X:


4K, DLSS Quality Transformers, MFG 4X:


The key thing from those three images is the FPS. Note also that the frametimes chart is fubar because RTSS / CapFrameX doesn't know how to deal with the new MFG pacing algorithm (yet). So in this case, MFG is providing nearly perfect scaling between 2X and 4X — 106 vs 212 FPS. Which means the input is still at the same rate (53 FPS) but you're getting more frames. There's a bit of wiggle room on the 3X result, just because this isn't a full benchmark. But I do have full benchmarks! For only the 5090...

CP77 FullRT 4K Native             AVG:  29.78   1%Low:  24.2
CP77 FullRT 4K DLAA               AVG:  30.38   1%Low:  23.9
CP77 FullRT 4K DLSSQ-CNN          AVG:  61.82   1%Low:  45.4
CP77 FullRT 4K DLSSQ-Transformers AVG:  58.15   1%Low:  42.8
CP77 FullRT 4K DLSSQT+FG2X        AVG: 106.88   1%Low:  73.7
CP77 FullRT 4K DLSSQT+FG3X        AVG: 153.75   1%Low:  93.7
CP77 FullRT 4K DLSSQT+FG4X        AVG: 195.67   1%Low: 104.1

So, interesting that DLAA actually outperformed native, probably because native doesn't use ray reconstruction and the denoisers for standard CP77 rendering are more complex than RR. The performance hit from transformers vs CNN looks to be relatively minor, about 6% (at least on the 5090).

Framegen gave an 83% boost to FPS with 2X mode, 164% with 3X mode, and 236% with 4X mode. Input framerate is 58.2 FPS without framegen, 53.4 FPS with 2X mode, 51.3 FPS with 3X mode, and 48.9 FPS at 4K mode. I dare say if that holds, people won't notice the minor change in input sampling rate and higher MFG levels will look smoother.

But that's really the crux of the story: What happens if you're not getting 50+ FPS on the input rate with framegen? Or more like what happens when the input sampling is only 25 FPS? In most games that I've tried with framegen (DLSS3 or FSR3), if you're getting a generated framerate of 50 FPS, the game feels very laggy. So I'd expect a similar situation with MFG running at 4X mode and 100 FPS.

 

[JarredWaltonGPU]

Something I haven't seen mentioned yet is the fact that current-gen gaming CPUs may actually be holding back the RTX 5090, but not the RTX 4090, at least not in any noticeable way.

The TomsHardware GPU testing rig currently in use features an AMD 9800X3D, which is an excellent gaming CPU, but there’s an interesting point to consider when it comes to memory bandwidth. Based on AIDA64 benchmarks, with DDR5 6000 dual-channel RAM, the 9800X3D shows a read speed of around 62 GB/s, a write speed of 84 GB/s, and a copy speed of 59 GB/s.

For PCIe 4.0 x16, which maxes out at 32 GB/s unidirectional and 64 GB/s bi-directional, this isn’t really an issue. The available system memory bandwidth is comfortably above what the PCIe bus can handle, meaning there’s little to no risk of a bottleneck between the PCIe controller and the system memory. The CPU and memory can keep up with the data demands without any major slowdowns.

However, things start to look a little different when you consider PCIe 5.0 x16, which doubles the bandwidth to 64 GB/s unidirectional and 128 GB/s bi-directional. With PCIe 5.0 pushing such high data rates, there's a possibility that the memory subsystem may become a limiting factor.

If the GPU is trying to fully saturate the PCIe 5.0 link, the RAM's read and copy speeds might struggle to keep up, potentially causing a bottleneck between the PCIe controller and the system memory. This could lead to situations where the 5090 isn't able to fully stretch its legs, particularly in scenarios involving high data throughput.

In contrast, the 4090, running on PCIe 4.0, wouldn't face the same challenge because the available system memory bandwidth is nearly equal to what the GPU can demand via the PCIe bus.

What do you all think? Could the memory bandwidth limitations of the 9800X3D be a contributing factor here? Do you think that the 5090 might eke out a few more points if paired with a system with higher RAM bandwidths? Would love to hear others' thoughts on this.

Someone else has already tested the 5090 running in PCIe 4.0, 3.0, 2.0, and 1.0 speeds. Basically, anything above 3.0 was fine. 3.0 was a couple of percent slower, 4.0 was the same performance, 2.0 was maybe 6% slower (?). I think that was the basic gist. So PCIe bandwidth isn't a factor for games. No surprise there as you have 1.8 TB/s of bandwidth vs up to 64 GB/s unidirectional. GPUs attempt to do as much as possible internally without worrying about transfers over PCIe. I think mostly it's geometry updates that use PCIe bandwidth while gaming.

 

[SydneyAngel]

Nvidia is acting like intel was back like 10 years ago. They know they have us. Its just business i guess

 

[Gaidax]

CP77 FullRT 4K Native             AVG:  29.78   1%Low:  24.2
CP77 FullRT 4K DLAA               AVG:  30.38   1%Low:  23.9
CP77 FullRT 4K DLSSQ-CNN          AVG:  61.82   1%Low:  45.4
CP77 FullRT 4K DLSSQ-Transformers AVG:  58.15   1%Low:  42.8
CP77 FullRT 4K DLSSQT+FG2X        AVG: 106.88   1%Low:  73.7
CP77 FullRT 4K DLSSQT+FG3X        AVG: 153.75   1%Low:  93.7
CP77 FullRT 4K DLSSQT+FG4X        AVG: 195.67   1%Low: 104.1

What would be interesting here is to see what happens with DLSS at Balanced and Performance modes, specifically for the DLSSQ-Transformers base line (58 FPS one).

I wonder how much FPS is gained by running "Balanced" preset VS the artifacts/quality degradation. Maybe the latter is near indistinguishable, while you gain 10-15 more FPS, which would make Framegen much better.

It would especially be interesting to see for card like 5080/5070Ti, because that might be all that is needed to make Framegen usable. As you write, it needs some basic decent FPS to begin with to work decently well.

 

[M0rtis]

You know what I think ?

I think that the 5090 should be irrelevant from our (gamer) POV
The top of the line card that we should all want or be interested in is the 5080.
The xx80 series was always the top of the line card for gamers.
Somewhere along the line, I think with the 3000 series, they changed the nomenclature of the Titan cars to xx90 cards. Overnight, thanks to the magic of marketing we automatically switched form thinking that the Titan cards, which were outliers and meant for borderline professional use, benchmark flexing, nerd shenanigans and so on, were part of the same series of cards as the regular ones because of how they were labelled.
Obviously the crypto boom and the high priced scalped cards led to major FOMO, hey, its human nature and we began to consider Titan cards to be the top tier gaming cards rather than the xx80 series instead of the "Halo" cards that they were meant to be.
Its also thanks to poor game optimization and the jump to 2k and 4k as well as people being made to think that 100 FPS is the bare minimum framerate below which games are absolutely unplayable, that has reinforced the 90 series as the top tier card.

I dont have any numbers on hand and the benchmarks for the 5080 arent out yet, but can anyone tell me if the 80 series has gotten a nerf since the 2000 series and has the performance gap between the 80s and 90s widened ?

I hate to be that guy but Wake up Sheeple !! We should be demanding more of the 80 series cards and they should be the ones in the limelight and what we as gamers want for our next high end rig. Obviously its not going to happen now because of how people are so it will henceforth continue to be relegated to the "upper mid range"

 

[thestryker]

Something I haven't seen mentioned yet is the fact that current-gen gaming CPUs may actually be holding back the RTX 5090, but not the RTX 4090, at least not in any noticeable way.

The TomsHardware GPU testing rig currently in use features an AMD 9800X3D, which is an excellent gaming CPU, but there’s an interesting point to consider when it comes to memory bandwidth. Based on AIDA64 benchmarks, with DDR5 6000 dual-channel RAM, the 9800X3D shows a read speed of around 62 GB/s, a write speed of 84 GB/s, and a copy speed of 59 GB/s.

For PCIe 4.0 x16, which maxes out at 32 GB/s unidirectional and 64 GB/s bi-directional, this isn’t really an issue. The available system memory bandwidth is comfortably above what the PCIe bus can handle, meaning there’s little to no risk of a bottleneck between the PCIe controller and the system memory. The CPU and memory can keep up with the data demands without any major slowdowns.

However, things start to look a little different when you consider PCIe 5.0 x16, which doubles the bandwidth to 64 GB/s unidirectional and 128 GB/s bi-directional. With PCIe 5.0 pushing such high data rates, there's a possibility that the memory subsystem may become a limiting factor.

If the GPU is trying to fully saturate the PCIe 5.0 link, the RAM's read and copy speeds might struggle to keep up, potentially causing a bottleneck between the PCIe controller and the system memory. This could lead to situations where the 5090 isn't able to fully stretch its legs, particularly in scenarios involving high data throughput.

In contrast, the 4090, running on PCIe 4.0, wouldn't face the same challenge because the available system memory bandwidth is nearly equal to what the GPU can demand via the PCIe bus.

What do you all think? Could the memory bandwidth limitations of the 9800X3D be a contributing factor here? Do you think that the 5090 might eke out a few more points if paired with a system with higher RAM bandwidths? Would love to hear others' thoughts on this.

While I doubt this is an issue it would be an interesting test to run. This would realistically require testing on ARL though. The tests would need to be run in Gear 2 with 6000 CL30 and 8800 CL42 (or 8000 CL40, but ideally higher speed) to keep the same latency. Then compare 5090 and 4090 results to see if there's any uplift difference.

 

[spongiemaster]

That's exactly what I don't agree on your arguement, you bought the halo CPU which could max out your next few halo GPU purchase at 4k, that's where the point you buy a halo product, to stay at the top attainable performance longer. Yes the 9800X3D is expensive now, but you can be assured at least in 6090 or even 8090 it won't be bottlenecking your rig in 4k gaming. But for this Halo GPU? you actually dont even able to max out the mid range of last 2 gen of CPUs, that's called paying for bottlenecking at Halo usecase. That DLSS frame gen is never a true HALO feature, yes it gens a ton of frames to smooth out your graphics, but there are glitches here and there where you could notice in games during sessions, it isn't always popping up, but paying $2000 for a card where at the year of release required those frame gen to play new titles smoothly at 4k, while in the meantime having annoying glitches (e.g. starwars outlaw in some youtube reviews having weird textures now and then, flight sime altimeter ghosting)


And those ppl don't even buy a GPU let alone the 5090...

In the US, the 9800X3D isn't available for purchase rendering your entire argument, especially anything related to price, moot. The 7800X3D is currently selling for $480 from major 1st party vendors. That's $30 more than the launch MSRP 2 years ago. Where are the accusations that AMD is intentionally limiting supply to drive up demand and prices? If you're predicting price drops below MSRP for the unavailable 9800X3D any time soon, good luck with that.

What settings do we use to get a 9800X3D to more than double the performance of a 5800X3D on average like an RTX 5090 does vs a 3090?

"Futureproofing" is one of the dumbest things any tech person can use to justify any purchase. A 7600X is 3% slower than a 9800X3D at 4k and can be bought for 40% of what a 7800X3D sells for now. 4 years from now, you'll be able to use the $300 you saved today to buy a CPU that will crush the 9800X3D by 3 or 4% at 4k.

 

[Peksha]

Ada Lovelace and Blackwell architectures are distinctly different.

What do you mean by significant? List more than one difference compared to Ada?

 

[steven4570]

Comparing a GPU to a Rolex is RIDICULOUS. I can get $20 watch that does the EXACT same thing as a Rolex. A Rolex is a luxury product. A GPU isn

Comparing a GPU to a Rolex is RIDICULOUS. I can get $20 watch that does the EXACT same thing as a Rolex. A Rolex is a luxury product. A GPU isn’t.

It's really not. A 5090 is a luxury product. If you think it's not, then maybe you need to go outside and get some perspective.

 

[steven4570]

Hence why that entire comment was prefaced with the statement "personally". You can rate it however you would like, you also glossed over the other reasons, but to each their own.

There really are no other real reasons. Saying a 5090 is not faster the faster card is about as dumb as a gets.

 

[steven4570]

yes it does when it costs more than what it gives.

it's the fastest video card on the planet. You can hate on the price all you want, I'll even join you to a certain degree. But a 2.5 star product implies that if you buy this and use it's a mediocre and honestly bad experience. And it won't be, no one would say that. You can't call the fastest and best performing card mediocre and be taken seriously.

 

[Thunder64]

it's the fastest video card on the planet. You can hate on the price all you want, I'll even join you to a certain degree. But a 2.5 star product implies that if you buy this and use it's a mediocre and honestly bad experience. And it won't be, no one would say that. You can't call the fastest and best performing card mediocre and be taken seriously.

The 4090 was a much better upgrade and it got 4.5 stars. Therefore, the 5090 should be lower. I think 3.5-4 is fair.

 

[helper800]

What do you mean by significant? List more than one difference compared to Ada?

Ray Tracing Cores: 4th Gen
Tensor Cores: 5th Gen
NVENC: 9th Gen
NVDEC: 6th Gen
PureVideo HD: VP13
VDPAU: Feature Set M

Vs

Ray Tracing Cores: 3rd Gen
Tensor Cores: 4th Gen
NVENC: 8th Gen
NVDEC: 5th Gen
PureVideo HD: VP12
VDPAU: Feature Set L

4NP node (only for Blackwell GB100) vs 4N node.
GDDR7 (HBM3e in the case of the B200) vs GDDR6X.

Please check the links I posted, there are many more differences than the above.

 

[newtechldtech]

Ray Tracing Cores: 4th Gen
Tensor Cores: 5th Gen
NVENC: 9th Gen
NVDEC: 6th Gen
PureVideo HD: VP13
VDPAU: Feature Set M

Vs

Ray Tracing Cores: 3rd Gen
Tensor Cores: 4th Gen
NVENC: 8th Gen
NVDEC: 5th Gen
PureVideo HD: VP12
VDPAU: Feature Set L

4NP node vs 4N node.
GDDR7 vs GDDR6X.

Please check the links I posted, there are many more differences than the above.

does not translate into performance ... benchmarks it what counts !

 

[Peksha]

But a 2.5 star product implies that if you buy this and use it's a mediocre and honestly bad experience. And it won't be, no one would say that. You can't call the fastest and best performing card mediocre and be taken seriously.

We can and will)
The value proposition of the product is assessed. As a next-gen product, it's a failure. These outdated SMs got no love from NV, and are unable to use all those shiny new things - even the 78% increased bw VRam or pcie5 mentioned in the pros. But no one mentioned the zero changes in the main part of the SMs architecture, as well as the artificial supply shortage in the cons. Right now, the value of the 4090 is superior to this new 5090 - and the only recommendation should be - "forget about it and save your money." Naturally, the sponsor of the review does not approve of this. But if a review requires sponsor approval, then this is already marketing material, not a review.

"Ray Tracing Cores: 4th Gen
Tensor Cores: 5th Gen
NVENC: 9th Gen
NVDEC: 6th Gen
PureVideo HD: VP13
VDPAU: Feature Set M
"+more etc:
this doesn't use for rendering.

 

[Gaidax]

We can and will)
The value proposition of the product is assessed. As a next-gen product, it's a failure. These outdated SMs got no love from NV, and are unable to use all those shiny new things - even the 78% increased bw VRam or pcie5 mentioned in the pros. But no one mentioned the zero changes in the main part of the SMs architecture, as well as the artificial supply shortage in the cons. Right now, the value of the 4090 is superior to this new 5090 - and the only recommendation should be - "forget about it and save your money." Naturally, the sponsor of the review does not approve of this. But if a review requires sponsor approval, then this is already marketing material, not a review.

"Ray Tracing Cores: 4th Gen
Tensor Cores: 5th Gen
NVENC: 9th Gen
NVDEC: 6th Gen
PureVideo HD: VP13
VDPAU: Feature Set M
"+more etc:
this doesn't use for rendering.

You say that, but can you truly say that none of it is used for rendering when you finally have games that outright require Ray Tracing like Indiana Jones and now upcoming Doom?

Or can you say that Tensor Cores are not used in rendering when the new DLSS is spectacular, or if you're a purist you can use DLAA instead which is provides stellar antialiasing without sitting on your raster?

I think it's a tad dishonest there.

You could say that standard shaders did not change sufficiently, but there were changes there too integrating neural processing support in hardware with upcoming DirectX version exposing this to game devs via standard API.

I think some people have a real hard time recognizing that times change. Your final output on your screen is not only good 'ol raster nowadays at highest fidelity levels for latest most visually intense titles. RT starts to matter a lot now.

 

[UnforcedERROR]

It's really not. A 5090 is a luxury product. If you think it's not, then maybe you need to go outside and get some perspective.

Part of the appeal of a Rolex is that they retain their value. They're luxury, but they're often also a financial investment. It's a bad comparison in this respect. I suspect it's why they didn't take issue with the luxury car part of your statement, since most cars lose value.

it's the fastest video card on the planet. You can hate on the price all you want, I'll even join you to a certain degree. But a 2.5 star product implies that if you buy this and use it's a mediocre and honestly bad experience. And it won't be, no one would say that. You can't call the fastest and best performing card mediocre and be taken seriously.

There's more to a product evaluation than just "fastest best." I'm not weighing into the argument of what I believe it should be rated, but I understand why people think 4.5 feels high. I think it's fair to say the top-end of a spectrum can also be a poor product relative to many factors, especially those with a better value proposition. The 5090 simply isn't a revolutionary product, I think we can all agree on that.

does not translate into performance ... benchmarks it what counts !

But that's not the issue he took with your statement. The performance wasn't the problem, it's that you implied it was the same architecture getting those results. It's not, period.

 

[Peksha]

So PCIe bandwidth isn't a factor for games.

This is not true. When you reduce the resolution or use strong scaling, it starts to have a strong impact on performance. For example, the ultrafast performance preset can encounter exactly this problem. All eGPU users also know about this.

It doesn't help that this video card can't use the full bandwidth of pcie 5.0x16, despite the sticker on the box (well, you can run aida gpgpu or c&c test to confirm or deny this, but until you do we see that nothing is happening).

All this will play a role when this GPU is running in pcie x8 mode (Intel B690-B890/AM5 B650-870 with m2) in ultraperformance mode for PT/RT/etc

 

[Peksha]

You say that, but can you truly say that none of it is used for rendering when you finally have games that outright require Ray Tracing like Indiana Jones and now upcoming Doom?

If hw-RT is important for you, did you find any improvements in RT performance in the test results, or is this just marketing hype?
Also all these games work fine without hw/rt.
Are there any other announced games on the over popular Id Tech 7 engine?
Did you know that RT calculations don't require specialized hardware, so most RT games simply don't use your shiny HW-RT cores (along with all their shiny parts), and won't for a long time?

 

[Stomx]

Pity greedy NVIDIA removed NVlink. With NVlink I would just buy some used 3090Ti for $800-1000 in addition to my existing card and got the same performance as 5090 if not more because of larger VRAM 48GB vs 32GB in 5090

 

[YSCCC]

In the US, the 9800X3D isn't available for purchase rendering your entire argument, especially anything related to price, moot. The 7800X3D is currently selling for $480 from major 1st party vendors. That's $30 more than the launch MSRP 2 years ago. Where are the accusations that AMD is intentionally limiting supply to drive up demand and prices? If you're predicting price drops below MSRP for the unavailable 9800X3D any time soon, good luck with that.

What settings do we use to get a 9800X3D to more than double the performance of a 5800X3D on average like an RTX 5090 does vs a 3090?

"Futureproofing" is one of the dumbest things any tech person can use to justify any purchase. A 7600X is 3% slower than a 9800X3D at 4k and can be bought for 40% of what a 7800X3D sells for now. 4 years from now, you'll be able to use the $300 you saved today to buy a CPU that will crush the 9800X3D by 3 or 4% at 4k.

Speaking of supply and inflated price, NVIDIA is the worst offender of the whole world since the 3000 series… the price tripled from MSRP.

And so what do you gain going from a 4090 to a 5090? Without dlss frame gen they don’t chew threw Ray tracing games well at 4K, with frame gen artefacts. So in this sense the 5090 is, paying for “future proof”

And no, saving $300 from a CPU and you can’t buy something as good in three years with the $300 compared to getting a 7800X3D now. In three years upgrading a cpu will require a new motherboard and ram at least, while GPU upgrading is painless, one could easily going from a 1080Ti to 5090 with the totl platform without bottlenecking the 5090 much

 

[newtechldtech]

But that's not the issue he took with your statement. The performance wasn't the problem, it's that you implied it was the same architecture getting those results. It's not, period.

I was talking about the performance. Period ! actually it does not seem as a new design at all , just renaming and tweaks . Period.

 

[steven4570]

We can and will)
The value proposition of the product is assessed. As a next-gen product, it's a failure. These outdated SMs got no love from NV, and are unable to use all those shiny new things - even the 78% increased bw VRam or pcie5 mentioned in the pros. But no one mentioned the zero changes in the main part of the SMs architecture, as well as the artificial supply shortage in the cons. Right now, the value of the 4090 is superior to this new 5090 - and the only recommendation should be - "forget about it and save your money." Naturally, the sponsor of the review does not approve of this. But if a review requires sponsor approval, then this is already marketing material, not a review.

"Ray Tracing Cores: 4th Gen
Tensor Cores: 5th Gen
NVENC: 9th Gen
NVDEC: 6th Gen
PureVideo HD: VP13
VDPAU: Feature Set M
"+more etc:
this doesn't use for rendering.

You can and won't be taken seriously then.

 

[steven4570]

Part of the appeal of a Rolex is that they retain their value. They're luxury, but they're often also a financial investment. It's a bad comparison in this respect. I suspect it's why they didn't take issue with the luxury car part of your statement, since most cars lose value.


There's more to a product evaluation than just "fastest best." I'm not weighing into the argument of what I believe it should be rated, but I understand why people think 4.5 feels high. I think it's fair to say the top-end of a spectrum can also be a poor product relative to many factors, especially those with a better value proposition. The 5090 simply isn't a revolutionary product, I think we can all agree on that.


But that's not the issue he took with your statement. The performance wasn't the problem, it's that you implied it was the same architecture getting those results. It's not, period.

So it has to retain value in order to be a luxury item? LOL The mental gymnastics are incredible. I Never said a 5090 was a revolutionary product.
I also Never implied it was the same architecture.