Nvidia's 3000 Series RTX GPU [3090s with different memory capacity]

Pascal went that low. The GT 1030 is still available today for $85.
That GP108 is 3 years old, though.

I think Renoir's high-clocked Vega the Tiger Lake's Xe have practically killed that performance range for dGPUs.
Intel did release the Xe MAX dGPU, supposedly to bundle with the older 14nm++(...)+ laptop CPUs with gen9 graphics, but I haven't seen much of an adoption there.
I wonder if they initially planned for this GPU to work in parallel with TigerLake graphics on DX12 explicit multiadapter, but then the latter never got any kind of wide adoption.


Regardless, iGPUs in later APUs/SoCs seem to have killed the $100 value proposition in dGPUs. I can only guess that DDR5 and Infinity Cache on AMD APUs are bound to make that even more clear.
 
Agreed. I do hope Intel stops their weird practice of lumping their highest-end IGPUs only with their highest-end CPUs. Anybody spending cash on a high-end CPU who also cares about GPU performance will definitely buy a discrete GPU, so the IGPU is just dead weight. On the other hand, attaching a top-end IGPU with a low-end or midrange CPU can make for some really compelling efficient SFF builds.
 
That GP108 is 3 years old, though.

I think Renoir's high-clocked Vega the Tiger Lake's Xe have practically killed that performance range for dGPUs.
Intel did release the Xe MAX dGPU, supposedly to bundle with the older 14nm++(...)+ laptop CPUs with gen9 graphics, but I haven't seen much of an adoption there.
I wonder if they initially planned for this GPU to work in parallel with TigerLake graphics on DX12 explicit multiadapter, but then the latter never got any kind of wide adoption.


Regardless, iGPUs in later APUs/SoCs seem to have killed the $100 value proposition in dGPUs. I can only guess that DDR5 and Infinity Cache on AMD APUs are bound to make that even more clear.

It’s not just about performance. People don’t upgrade their CPUs that often. GPU upgrade cycles are shorter so you have more opportunity to adopt new features.
 
Yes you love Ampere and the 3060 and how much it would hypothetically beat the PS5 in a hypothetical RT-heavy scenario that will never happen because games made for the PS5 won't run RT-heavy scenarios that only run decently on RTX30 hardware.
My bad, probably I formulated my previous message too poorly for you, probably too many technical terms, let me put it simple for you in the "love/hate" speech you understand.
Yes, I know you hate everything NVIDIA and will try to make it look bad and evil whenever you can, but I don't really care.
What I care about though is the nonsense about the performance-per-TFLOP metric. As I noted before, this metric is useless in separation from other metrics, so you can't make some stupid claims based on 6800 XT vs 3080 FLOPS comparisons.
Probably you didn't understand this from my first attempt because I had mentioned both NVIDIA and AMD. Let boil it down just to AMD, so it's easier for you to understand - PS5 is NOT 20% slower than XSX and PS5 doesn't have a significantly higher performance-per-TFLOP, it's basically the same architecture all around.
Both consoles have different strengths - higher triangle setup rate, fill rate and blending for PS5 and higher TFLOPs and texture fetch/filtering for XSX, this results in both having almost identical avarage performance in current rasterisation games despite of 20% FLOPS shortage on PS5 side.
Future engines, such as UE5, might move the needle to the compute side of things (it has been the case for last 10 years), chips with higher TFLOPs will benefit the most from increasing the amount of computations and that's already the case with RT titles like it or not.
 
Last edited:
My bad, probably I formulated my previous message too poorly for you, probably too many technical terms, let me put it simple for you in the "love/hate" speech you understand.
Yes, tell me more about your subpixel rendering theories.


Yes, I know you hate everything NVIDIA and will try to make it look bad and evil whenever you can, but I don't really care.
You really don't care at all, which is why you're totally not coming here guns-blazing to avenge one certain dear IHV at every passing hint of slight criticism.


Probably you didn't understand this from my first attempt because I had mentioned both NVIDIA and AMD. Let boil it down just to AMD, so it's easier for you to understand - PS5 is NOT 20% slower than XSX and PS5 doesn't have a significantly higher performance-per-TFLOP, it's basically the same architecture all around.
Ever heard of linear equations in school? Math?

If the PS5 is not 20% slower than the SeriesX while having 20% less TFLOPs, then it has a higher performance-per-TFLOP. If rendering-performance-PS5 with 1T compute throughput is similar to rendering-performance-SeriesX with 1.2T compute throughput, then the PS5 has higher rendering-performance-per-TFLOP.

I can direct you to some youtube videos about linear equations if you want. I'm here to help!


Both consoles have different strengths - higher triangle setup rate, fill rate and blending for PS5 and higher TFLOPs and texture fetch/filtering for XSX, this results in both having almost identical avarage performance in current rasterisation games despite of 20% FLOPS shortage on PS5 side.
Future engines, such as UE5, might move the needle to the compute side of things (it has been the case for last 10 years), chips with higher TFLOPs will benefit the most from increasing the amount of computations and that's already the case with RT titles like it or not.
Nah, your previous post clearly said "b-but the RTX3060 is probably faster than the PS5 at RT-heavy games", which is still a bullshit comparison because can never be measured by the general public. Just like nvidia's previous claims of the RTX2080 Super laptops with a performance equivalent to a GTX1070 being faster than both the PS5 and the SeriesX.
You can try to change goalposts and narratives to fit a completely different argument, but anyone can follow links and see what is written in the post I answered to.
 
Last edited by a moderator:
Ever heard of linear equations in school? Math?
I've used it many times for performance modelling at different scales and levels, did you?;)
Have you heard of linear regression?
Try it and then hopefully you won't repeat the nonsense with performance-per-TFLOP.

If the PS5 is not 20% slower than the SeriesX while having 20% less TFLOPs, then it has a higher performance-per-TFLOP
As I already said, performance in rasterisation games is a mix of shaders, rasterization and other workloads, i.e. there is a number of variables, not just a single variable, and you can get the right solution by making a system of linear equations and optimizing regressors (evolutionary or by SGD).
Regressors here will be an average percentage of time spent in different workloads which rasterization consists of. SeriesX can't have less performance-per-TFLOP because it doesn't hold true for compute workloads where it will be proportionally faster. As we know there are such games as Clay Book and Dreams which are almost 100% made of compute shaders.
If you think of a complex workload, such as rasterisation, as of a simple linear equation with a single unknown then you are doing it wrong.

which is still a bullshit comparison because can never be measured by the general public
Well, I can clearly see you are unaware of perfomance modelling.
With RDNA2 performance data, it's quite simple to model how PS5-specced hardware will hold up in any game.
I agree only on that's not a very correct comparison because PS5 has its own software stack which is another part of equation and it's hard to model unless you are a PS5 developer with devkits.
If they wrote PS5 level PC hardware, this would have been other story.
 
Last edited:
Visualizing GL_NV_shader_sm_builtins – Wunk (wunkolo.github.io)

GL_NV_shader_sm_builtins is no longer in the Nvidia beta-driver-jail, and can allow shaders to access the index of the physical Streaming multi-processor and warp that a particular compute kernel dispatch is executing upon.
...
So, on my GTX 1650, we’re expecting 16 Streaming-Multiprocessors, and each SM to have 32 warps per SM. And now with this extension we have a means to identify both the index of the SM([0,15]) and the identifier of the Warp executing within that SM([0,31]) within my particular GPU.
TU117-grid.gif
 
Geez there is way to much emotion in some of these posts.

RDNA2 has a significantly higher performance-per-TFLOP than Ampere on rasterization, considering the 30 TFLOPs RTX3080 competes with the 20TFLOPs 6800XT.

3The RTX 3 is only true when no INT execution is happening anywhere on any SM (otherwise FP32 throughput for that SM is halved). Yes, you can say you are just quoting Nvidia's marketing numbers, but if you're actually interested in truth, then the RTX 3008 is not a 30 TFLOP GPU. AMD/Nvidia TFLOP comparisons are no longer straightforward. Seeing as this is a technical forum I think we can probably look past marketing
 
December Steam hardware survey shows 3080 at 0.48%, which is a very fast 2-month ramp. Wish I had historical data to compare vs. prior $700 GPU launches, but comparing that number with other current numbers, e.g., 1650Ti (a 2019 GPU with both desktop and mobile parts) at 0.57% and 5700XT at 0.89% tells me that there are a large number of 3080s in the hands of gamers. Maybe they were all bought from scalpers, I don't know, but the numbers are high. There's no question that both Nvidia and AMD have supply problems, and demand is so high that to end users hitting F5 it seems like the cards don't exist. Sucks to be us, sucks to be them.

On a side note, I wonder what happens towards the latter half of the year once the vaccines kick in and everyone is up and about partying like its 2019 all over again. Do we see a huge crash in the at-home entertainment business? A glut of chip inventory?

If you go to the old Steam Survey thread you can find little tidbits in there. For example:

https://forum.beyond3d.com/posts/1271140/

Radeon HD 6850 after a little over a month sat at 0.49% while GTX 580 was at 0.55% after about a month. There's no number reported for Radeon HD 6870, but one poster commented that it was higher than either of those 2 cards.

I also commented later in the thread that it was interesting that Radeon HD 6870 was performing better in the Steam survey than the Radeon HD 6850. I found that interesting as I expected the cheaper Radeon 6850 to have a higher share of Steam Users than the 6870.

I don't feel like digging through that entire thread, but IIRC both 4870 and 5870 performed even better than 6870 in Steam Survey when they launched, unfortunately at the time Steam grouped all Radeon 58xx users into one number, so it's difficult to determine how many were 5870 and how many were 5850. But combined, they were greater that 6870 and 6850, IIRC.

At least WRT Steam Survey, the RTX 3080 isn't anything special in terms of hardware install and is behind quite a few other historical cards when they debuted in Steam Survey. This is especially true considering that it took the 3080 2 months to get to that point where many launch cards got to that point in a month or less.

However, various factors don't make these numbers directly comparable. So, while interesting, there isn't much real information we can glean from it. Also, if it's true that many RTX 3xxx cards ended up with Cryptominers, then that means less getting into actual gaming rigs with Steam installed.

Regards,
SB
 
Last edited:
3The RTX 3 is only true when no INT execution is happening anywhere on any SM (otherwise FP32 throughput for that SM is halved). Yes, you can say you are just quoting Nvidia's marketing numbers, but if you're actually interested in truth, then the RTX 3008 is not a 30 TFLOP GPU. AMD/Nvidia TFLOP comparisons are no longer straightforward. Seeing as this is a technical forum I think we can probably look past marketing
And what's happening on RDNA2 when there's the same INT execution happening anywhere on any WGP?
 
If you go to the old Steam Survey thread you can find little tidbits in there. For example:

https://forum.beyond3d.com/posts/1271140/

Radeon HD 6850 after a little over a month sat at 0.49% while GTX 580 was at 0.55% after about a month. There's no number reported for Radeon HD 6870, but one poster commented that it was higher than either of those 2 cards.

I also commented later in the thread that it was interesting that Radeon HD 6870 was performing better in the Steam survey than the Radeon HD 6850. I found that interesting as I expected the cheaper Radeon 6850 to have a higher share of Steam Users than the 6870.

I don't feel like digging through that entire thread, but IIRC both 4870 and 5870 performed even better than 6870 in Steam Survey when they launched, unfortunately at the time Steam grouped all Radeon 58xx users into one number, so it's difficult to determine how many were 5870 and how many were 5850. But combined, they were greater that 6870 and 6850, IIRC.

At least WRT Steam Survey, the RTX 3080 isn't anything special in terms of hardware install and is behind quite a few other historical cards when they debuted in Steam Survey. This is especially true considering that it took the 3080 2 months to get to that point where many launch cards got to that point in a month or less.

However, various factors don't make these numbers directly comparable. So, while interesting, there isn't much real information we can glean from it. Also, if it's true that many RTX 3xxx cards ended up with Cryptominers, then that means less getting into actual gaming rigs with Steam installed.

Regards,
SB
Thanks for digging up the data. I wasn’t trying to claim that the 3080 sold more/faster than those older cards. I was arguing that it sold enough to dislodge a visible fraction of the existing Steam GPU install base, which is already likely inflated due to the pandemic. But yeah it’s a tricky argument to make without absolute numbers.
 
Thanks for digging up the data. I wasn’t trying to claim that the 3080 sold more/faster than those older cards. I was arguing that it sold enough to dislodge a visible fraction of the existing Steam GPU install base, which is already likely inflated due to the pandemic. But yeah it’s a tricky argument to make without absolute numbers.

It will only start getting interesting when the 3060, 3070 and AMD's 6800 (hopefully differentiated) series numbers start to appear as you start getting an idea about the mix of cards.
 
INT and FP have alway competed for the same execution units on AMD and Nvidia hardware. Now that Nvidia is talking explicitly about shared FP+INT pipelines maybe folks think that INT is free on AMD. I guess we can blame Nvidia marketing for this too.
The only difference between Ampere and RDNA2 here worth discussing is the fact that RDNA2 can run INTs on all of its SIMDs giving it the same peak throughput as that of FP32 and Ampere is only able to run INTs on half of its SIMDs meaning that its INT throughput is half of its FP32.
 
I think sniffy was just trying to explain to ToTTenTranz why in some gaming scenarios the 20TFLOP RDNA2 is equal to the 30TFLOP Ampere and simply asked to look past the basic marketing numbers that most of us here don't adhere to anyway. We all know TFLOPS isn't everything.

I don't think there was a need to bring in the NDS.
 
Back
Top