Nvidia Ampere Discussion [2020-05-14]

How many billion triangles per second do we need with current-gen (i.e. the ones being benchmarked) games?

 

Their numbers are contradictory to other leaks, their TimeSpy Extreme results put the 3090 at only 10% faster than 3080, while the leaks show a 20% advantage.

https://videocardz.com/newz/nvidia-geforce-rtx-3090-3dmark-time-spy-scores-leaked

 

It's 7 GPCs against 6 GPCs (unless some RTX 3080's maybe have 7 GPCs?). All resources should have a >10% (really >20%) advantage with 3090 over 3080 with the exception of stock power limits. At least for the FE models that is 350w against 320w, which is only 9.4%, an issue that might be compounded by having 2x the memory chips.

 

This seems like a power limit to me. The 3090 with its gigantic cooler might be made for pushing way beyond the 350W TDP/TBP they're declaring.

I look at their coolers, with the 3090's being over 2x larger than the 3080's, and it doesn't make any sense that the former is only pushing 30W (10%) more than the later:





10% is the difference in performance we're seeing in these benchmarks, but it's also the difference in TBP. The 3090 really doesn't look like a 350W card.

Is there a standardized limit to push only up to 350W in a PCIe graphics card? If so, perhaps nVidia will be pushing for 3090 customers / reviewers to always overclock these cards as much as they can, and this is intended to be more like a 450-500W card than a 350W one.

 

How dare they.

 

I guess we know why nVidia is calling the 3080 their flagship card.

 

I want to see someone do a minor undervolt to remove power limiting and then try to push clocks as high as possible. Doesn't make sense to me that reviewers are pushing clocks when they're already power limited at stock.

 

It's the 3090 i want, or the 4090 at a later stage. Won't be having it this year but il watch others playing cyberpunk on that thing.

 

I am starting to think, that no matter what % we pin 3090's performance over the 3080, we are learning that Ampere doesn't scale well to Games. And I think that is why these cards are so aggressively priced, because RDNA is a more (game) focused architecture and supposedly scales well.

I've been building PC's since the "turbo button" days... and I can not think of a time, where I've anticipated so much in building a complete new rig, with a full blown dGPU.

The way I am seeing it, is NVidia being sort of a dual-purpose high-end solution, also being used in a semi-professional level for Content Creators and other Compute workloads, etc. Where as RDNA, is different from CDNA and will strictly be limited to Games & Frames..!

(?)

 

I think it's more likely these cards are built more for next-gen games.

 

I think he didn't think of what UE5 demo was doing. AMDs gpu's are probably build for the same purpose, gaming and professional market, just like they are now. Just be happy if they can match Turing/ampere performance. Were up to the point of close to 40TF of gpus, with 20 becoming the norm down the line. Progression isn't anything to complain about now in special DLSS and ray tracing advancements. Never thought a 3080 would outperform the 2080 by so much (between 80 and 100%), beating the 2080Ti handidly at a much lower price. And now AMD is coming. Good times.

 

More like it was the cheapest way to improve performance without making the architecture to wide. Doubling the FP32 throughput per SM is still a shocker.

 

I really think a lot of games are trapped in old-fashioned engines that don't scale well - some reviewers feature old games with "bad graphics" that just aren't scaling it seems. 4K resolution is practically the only data point in most games that isn't substantially "CPU-limited".

It would be nice if someone did some tests of scaling based upon "super-sampled 4K", i.e. 8K. But that's just going to show ROP/bandwidth advances in Ampere (which seem pretty decent). Super-sampling 1440p games that do not have a DLSS option is another test that it would be nice to see reviewers do. This would be very compelling for most enthusiast gamers who, it seems, have stayed at 1440p rather than suffer the pain of 4K gaming on prior GPUs.

We've seen in Gamers Nexus results that certain overclocking attempts badly hurt 1% lows (in some games, not all). This is perhaps a clue that power/thermals/"GDDR6X-fragility" are messing things up for Ampere. It could be nothing more than a symptom of the silicon lottery, teething problems that 6 months of manufacturing will ease.

Overall I think Ampere is merely showing that "eyecandy per GFLOP" is really poor for most games - many games will be kicking 3080 gamers in the teeth with their shitty image quality and DX9-dumb engines.

 

Seems logical NV/AMD design their hardware for future games, not older ones. I think those older games are not really left behind anyway, unless some aren't statisfied with 300+fps.

 

Although I generally agree, how are ROPs advanced in RTX 3080 compared to 2080 Ti? I mean apart from that paragraph in Nvidia's Whitepaper claiming them to be in a very doubtful way.

To me it looks like a very overpriced but potentially really quiet 350W card if you remove the useless two or three upmost freqency bins.

 

I presume you're referring to "[...] eliminating throughput mismatches between the scan conversion frontend and raster operations backend."

I think moving the ROPs inside the GPCs is probably a big deal. It changes the nature of the "crossbar" and associated queueing (which can cause back-pressure) between GPCs and memory channels, and I expect it will interact positively with tiled rasterisation (tiled "scan conversion").

Asus TUF appears to demonstrate that AIBs can make some very nice cards which are even quieter than FE. So maybe those frequency bins aren't so useless?

 

It's likely Ampere (and RDNA2) will still be the leading architectures for the next 2 years so they're primarily going to be judged on "next gen" game performance rather than current gen (which they'll both be more than capable of blitzing). Also, given the baseline for next gen console resolution over that time period is going to be 4k, a card like the 3080 will only really be judged at that resolution. If the XSX is running a game at 4k/60, who cares how fast the 3080 can run it at 1440p? It's 4k or bust. Therefore it makes sense that the GPU would be scaled to perform best at 4k at the expense of lower resolutions. It'll be interesting to see how the 3070 (my target card) compares across resolutions.

 

Yes, probably it will make things easier for Nvidia in the first place. But from what I've seen, ROP performance does not seem better than on 96-ROP-GPUs before. With rather edit: raster, I need to look into whether or not it's memory bandwidth bound.

Yes, the quiet mode on the Asus card is very nice. But it's already beyond 330 Watts and mostly using up a third slot - so, with the area increase for fins and fan blades on 3090, I think it can improve more compared to the 3080 FE than a 3090 TUF could compared to 3080 TUF.

 

3070 for 1440p gaming may well turn out dangerously close to 3080...

If you do some experiments, I suppose it would be worth adding super-sampled 1440p to your tests, since that's more pixels than 4K.

Though some games running at this resolution might exhaust the card's 10GB.

I was tempted to say this in my last post, but I've been badly burnt as an FE cooler fanboy...

 

Isn't this exactly what you would expect going from 320 to 350W power limit?