Nvidia Ampere Discussion [2020-05-14]

That's nice. Have they fulfilled launch day 3080 orders yet?

 

It wouldn't be the first time if that turns out to nothing, there has been tons of products registered at eec that never materialized (including just pure fake regististrations, someone registered geforce gtx 1180 etc under some video card brands name and got through to eec lists. the company had nothing to do with it. we've also had gtx 2070, 2080 etc)

 

TastyPC review. 380-400€. 2080+ performance. 40% more avg performance than the 2060 Super, 8GB VRAM, 200W, very close to the 3070 (which is almost like a 2080Ti)

 

the fact that bitcoin is again at record level worries me about these graphics cards quickly going out of stock 'cos of mining. The 3060Ti is the most efficient price/performance wise, gosh....-it's my favourite as of now and while I am saving I dunno how prices will fare-

 

It's neat to watch the 3060Ti stack up against the 3070 as it gives us some good indication as to which games are primarily bandwidth bound.
There are quite a few titles where the 3060Ti is only 1-2% slower than the 3070, and bandwidth is really the only explanation there, as all the other resources have been slashed by ~20%, including ROPs.

https://www.techspot.com/review/2155-geforce-rtx-3060-ti/

Metro:Exodus and Horizon Zero Dawn are two notable ones from the TechSpot review.

 

Some good news to see a entry level Ampere 3060Ti performing above a 2080 in raw performance. RT and DLSS will be even more impressive as compared to any Turing product probably.

 

Will it be even more impressive? From what I have got from the Ampere reviews so far is that although the new RT cores are better, there are less of them, and thus the percentage hit for enabling RT in games is similar to Turing?

 

No idea about 3060, but say someone going from a 2080 to a 3080 will see a substantial increase, sometimes even 200% in RT (Q2 RTX). I have a hard time believing a 3060Ti will perform worse then say a 2070/80 in RT, but we will see.

 

3060ti is similar perf on rt titles as 2080super fe is. Very decent 1440p card.

https://www.pcworld.com/article/359...-ti-founders-edition-review.html?page=5#toc-4

 

Thats impressive i think, a 2080Super was kinda the beast back then in all regards. Feels abit bad with the 2080Ti i have but il have to go with it for half a year or more, if i can sell it Over two years of service anyways.

 

So far it looks like the Ampere cards, even the GA104 based ones are just a bit better proportionally (IE, less percentage drop in performance over baseline) than their Turing counterparts with RT enabled.
Techpowerup has a decent comparison at the end of their 3060Ti reviews.

https://www.techpowerup.com/review/asus-geforce-rtx-3060-ti-strix-oc/36.html

 

This is another data point. For reference my 3070 renders that bmw scene with optix around 15.5s with slightly overclocked gpu and good overclock on ram. ray tracing improvement on ampere is bigger than rasterization performance improvement. How much bigger depends on how heavily rt is used and how well ray tracing is optimized for ampere.

6800xt renders this same scene somewhere around 38s at least based on benchmarks I have seen.



https://www.phoronix.com/scan.php?page=article&item=blender-290-rtx3080&num=3

 

Thanks for the link, that is very comprehensive. Gains in Metro: Exodus look better than I was expecting; @1080p Turing: 31-36% performance hit, Ampere: 26%. While gains for Control are less impressive; @1080p Turing: 43-47% performance hit, Ampere: 39-42%.

When thinking back to the Ampere lanuch the improvments nVidia talked about for their latest gen RT cores it's such a shame they scalled the quantity back.

Edit: Just to note that those Ampere numbers above are just for the 3060Ti and3070.

 

I'm going to have to dig to try to find the quote, but I remember reading back near the Turing release that Nvidia could have fairly easily doubled or even quadrupled the number of RT units if they wanted to, as they don't take a ton of die area. There wouldn't have been much point though, because when you have really heavy RT loads like Control with multiple features enabled, the limiting factor ends up being raw memory bandwidth more often than not; the working set is so large the small L2 cache (and apparently even AMD's giant infinity cache!) just can't keep up. It's why for GA102 in particular, the focus was on increasing memory bandwidth at all costs, as that was what would get them the RT performance they needed.

If GDDR6X costs come down, I wouldn't be surprised to see a GA104 refresh using it to boost RT performance. Nvidia explicitly said in their Ampere whitepaper that the GA104 memory controller is built to use either GDDR6 or GDDR6x. That extra bit of die area to make the memory controllers compatible with both types sits completely unused in all consumer GA104 products so far, but they knew it would be helpful to keep the future performance boost in their back pocket.

 

You do realize that you're just comparing 6800 XT, rendering it in generic OpenCL implementation which doesn't utilize RT acceleration, to NVIDIA cards running it on their very own and very optimized OptiX, which utilizes RT acceleration (when available) too, right?

RTX 3080 does it in about 26 seconds with CUDA, which like OpenCL has no RT acceleration, but is still very much NVIDIAs own and very well optimized for their products.
RX 6800 XT got 38 seconds in our test with OpenCL.

 

I realize that. It's however the best comparison we can do today or we have to refrain from doing any comparisons?

 

Here's a more recent, detailed Blender 2.90 review with all Ampere and Navi 2 cards evaluated together.
Blender 2.90: Best CPUs & GPUs For Rendering & Viewport – Techgage

 

Great to see that they're using larger tiles than the benchmark's respective defaults. This greatly helps the Radeon scores.

 

Looks like not even a 3090 boosted by DLSS can hit 60fps in Cyberpunk 4K maxed. Looking forward to DF’s in-depth image quality analysis because the visuals don’t seem that ground breaking.

These numbers are apparently from nvidia.

 

No need to be visualy ground breaking to have RT taxing a lot of perfs : /