Speculation: GPU Performance Comparisons of 2020 *Spawn*

[NightAntilli]

I think the real reason is that AMD has been caught with their pants down with the 3080 and 3070 performance and prices and that AMD will more than likely have to clock Big Navi well beyond the sweet power spot and pull a heck of a lot of watts to get near or slightly exceed the 3070 performance. So with this gear reviewer's can accurately measure performance per watt for both vendors.

Honestly, I hardly believe this to be the case, especially with cards like this...;

https://www.overclock3d.net/news/gp...90_xtreme_gpus_appear_to_be_4-slot_monsters/1

 

[Benetanegia]

Well, yeah, I realized something was wrong, but I decided to post it anyway, to see what the rebuttal would be. Helps me learn too.

That's good man.

I would be surprised if it exceeds 16.

Well like the quote from the blogpost says Doom 2016 uses 50 render targets, that means that the scene is rendered at least 50 times per frame in different ways, although not all of them are full res. But many of those are modified several times per frame so it's probably a much higher number.

But in any case, what I am really asking for is some evidence that the amount of GPixels/s you provided actually is a bottleneck.

That's really hard to do. in any case, I think it is more important to make it clear that it is not a bottleneck that they are talking about. Bottleneck implies that after a certain point increasing performance by increasing other units other than the bottleneck is imposible. That's not how GPUs work. Rendering a frame is divided into thousands of small semi-dependent tasks, some of which take place mostly on shaders, others mostly or exclusively on ROPs, texture units, etc. Increasing one particular type of unit will make its relevant tasks finish faster, thus reducing the frame time and increasing frame rate, although how much the fps will increase depends on how much of the frametime that particular unit has influence over.

 

[SimBy]

Honestly, I hardly believe this to be the case, especially with cards like this...;

https://www.overclock3d.net/news/gp...90_xtreme_gpus_appear_to_be_4-slot_monsters/1

I think it's pretty obvious at this point Nvidia marketing was very creative with those 1.9x perf/W improvement. If the 3080 benchmark leaks are the real deal it's only 30% faster than 2080Ti on average.

 

[pjbliverpool]

If the 3080 benchmark leaks are the real deal it's only 30% faster than 2080Ti on average.

Other than Ashes of the Singularity, what sources show this?

 

[Wesker]

I think the real reason is that AMD has been caught with their pants down with the 3080 and 3070 performance and prices and that AMD will more than likely have to clock Big Navi well beyond the sweet power spot and pull a heck of a lot of watts to get near or slightly exceed the 3070 performance. So with this gear reviewer's can accurately measure performance per watt for both vendors.

I’d buy this argument if it weren’t for the fact that the TBP of the RTX 3000 cards are well above 300 watts.

 

[CarstenS]

People need to stop staring at theoretical numbers that much, that 30 TF GPU isn't even twice as fast as same companys last gens 11 TF GPU

Do they, though? What if those people have workloads in mind, that rely on pure TFlops throughput? People should evaluate (i.e. stare at) things, they need or want from a given piece of hardware.

If AMD offers equivalent or better performance with lower power consumption it'll be a highlight of the reviews. I guarantee reviewers will mention that one requires a power supply upgrade and the other doesn't etc.

And they should, if that's the case.

 

[Kaotik]

Do they, though? What if those people have workloads in mind, that rely on pure TFlops throughput? People should evaluate (i.e. stare at) things, they need or want from a given piece of hardware.

Well, the so far leaked CUDA and OpenCL-benchmarks (CompuBench) aren't showing bigger gains than games, and they're mostly FP32 loads aren't they?

 

[CarstenS]

Well, the so far leaked CUDA and OpenCL-benchmarks (CompuBench) aren't showing bigger gains than games, and they're mostly FP32 loads aren't they?

I have no idea about the workloads from Compubench, your assumption would be as good (or better) as mine.

We have something from Nvidia directly though: Luxmark scales very well (>2x) apparently, Blender is also at or around 2x.
Of course, this has to be taken with an appropriately sized grain of salt. But I would wait until independent reviews pop up, leaks are IMO not very trustworthy, could be PR in one or the other directions, depending who got the hands on those cards.

 

[OlegSH]

Well like the quote from the blogpost says Doom 2016 uses 50 render targets, that means that the scene is rendered at least 50 times per frame in different ways

Yep, g-buffer fill, shadow maps fill, alpha blending, foliage overdraw, general overdraw w/o perfect front to back sorting, small triangles, etc, etc, all adds up to the ROPS usage.

 

[Geeforcer]

It's been very important while NVIDIA used less power than AMD equivalents

Not really. It has been important when the GPUs are close in performance, because then it becomes a differentiator. R290 series had a lot of power-related sins overlooked due to excellent performance, while various Fermi offshoots had it held against them even after Nvidia massacred their cost structure to make it quasi-competitive.

 

[NightAntilli]

Not really. It has been important when the GPUs are close in performance, because then it becomes a differentiator. R290 series had a lot of power-related sins overlooked due to excellent performance, while various Fermi offshoots had it held against them even after Nvidia massacred their cost structure to make it quasi-competitive.

And yet Fermi still sold like crazy and the R9 290(X) didn't.

 

[DegustatoR]

If the 3080 benchmark leaks are the real deal it's only 30% faster than 2080Ti on average.

On average of what? Fortnite and PUBG? Current gen games won't be able to make full use of Ampere h/w.

 

[DavidGraham]

And yet Fermi still sold like crazy and the R9 290(X) didn't.

Fermi was a good arch, it had immense performance, it's just that it's first iteration (GTX 400) was bad because of process problems, yet it still beat the HD 5870, NVIDIA fixed Fermi in the (GTX 500) series and still managed to beat the next AMD gen: HD 6970.

 

[Scott_Arm]

On average of what? Fortnite and PUBG? Current gen games won't be able to make full use of Ampere h/w.

Mesh shaders, variable rate shading, DXR 1.1 were already available in Turing. RTX IO will be available in Turing. Not sure what software developers would really have to do to leverage Ampere that would be different than Turing.

 

[Deleted member 2197]

Mesh shaders, variable rate shading, DXR 1.1 were already available in Turing. RTX IO will be available in Turing. Not sure what software developers would really have to do to leverage Ampere that would be different than Turing.

There also "motion blur hardware acceleration" to accelerate BVH ray tracing in high motion scenes which is unique to Ampere. Should be interesting once this is implemented in games.

 

[DegustatoR]

Mesh shaders, variable rate shading, DXR 1.1 were already available in Turing. RTX IO will be available in Turing. Not sure what software developers would really have to do to leverage Ampere that would be different than Turing.

There's like one game which use one of these features presently. But to take advantage of Ampere games have to push shading way harder than they do now.

 

[pjbliverpool]

There also "motion blur hardware acceleration" to accelerate BVH ray tracing in high motion scenes which is unique to Ampere. Should be interesting once this is implemented in games.

Also simultaneous use of FP32, RT and DLSS which games need to specifically support.

And of course generally higher RT loads.

 

[Scott_Arm]

Also simultaneous use of FP32, RT and DLSS which games need to specifically support.

Do they? Is simultaneous execution some extension of D3D? Extensions for Vulkan? Genuinely curious. I'd assumed this is something the driver would handle.

 

[T2098]

As he already told you in the very post you quoted, the 2070 has 48 rasterized pixels per clock, not 64. So actual relevant number of pixels are:

5700XT: 121.9 GPixel/s
2070S: 113.3 GPixel/s - limited by ROPs, rasterizers can actually do 141.6 GPixel/s because, 5 GPC x 16 (rasterized pixels) x 1770 Mhz

5700: 110.4 GPixel/s
2070: 77.76 GPixel/s - Limited by rasterizers: 3 GPC x 16 (rasterized pixels) x 1620 Mhz

Is this similar to the 1070 situation where it had 64 ROPS, but could only use all of them in some limited situations like MSAA resolve? (and in other situations it was limited by the front end to only 48 pixels/cycle)

 

[PSman1700]

Fermi was a good arch, it had immense performance, it's just that it's first iteration (GTX 400) was bad because of process problems, yet it still beat the HD 5870, NVIDIA fixed Fermi in the (GTX 500) series and still managed to beat the next AMD gen: HD 6970.

I had the GTX580 1.5gb, it was damn impressive for it's time, didn't get the 3gb since it wasn't available at the time. I had them in SLI for a while too, before going with a single GTX680.

There's like one game which use one of these features presently. But to take advantage of Ampere games have to push shading way harder than they do now.

Yes dont understand why people are concerned, these new gpus already perform like nothing else. It's the best thing that happen imo. Now AMD's turn.