Nvidia Ampere Discussion [2020-05-14]

I guess that might depend on how much you could win in the silicon lottery. I was under the impression, that things got a bit less variable after the introduction of finfet, but honestly haven't really made an effort to stay up-to-date with 7nm and beyond yet. Where there vastly different clock speeds in Navi10 with standard 7nm already?

And frankly, if 3090 already has 24 GiB G6X and since GA100 has no tensor cores, I could see this round going without a Titan-branded card. Just a few pro-features via driver for SPEC Viewperf is narrowing the niche even more.

 

That would make everything running on GPU software and hardware, you don't actually need tensor-cores for DLSS and tensor-cores aren't purpose built for DLSS

 

We've had the discussion about tensor core performance for AI reconstruction vs compute, even recently. I don't think we need to revisit yet again here.

 

that's the key -DLSS, RT, etc-.

I created an entry in the windows 10 calendar to notify me that the event started. I skipped through the past generation of cards but this one truly interests me

 

https://videocardz.com/press-release/asus-announces-geforce-rtx-30-series

 

We don’t know what we don’t know yet. It’s not like we are benchmarking release candidates. The nvidia RTX version of Minecraft could be very well optimized compared to the 4 week port DXR version of Minecraft.


Granted I still am on the same boat as you; but I don’t want to make a definitive statement yet on the whole of Navi 2 on an unoptimized product with a not mature devkit and possibly drivers. It is unfair to dog them before they had a chance to showcase.

 

I'm most interested to see what changes they will make to improve RT performance.

 

I am not basing this on the Minecraft demo alone, but also on the technical specs, RDNA 2 RT cores are shared twice, first with the Texture units for doing Intersections, and second with Shader cores for doing BVH traversal, it's a system optimized for die savings, NVIDIA's solution in Turing is fully independent and hence logically will be faster.

 

I think that you're oversimplifying something that's not so simple. We'll find out soon enough how the two compare.

 

A little under 100 minutes to go...

 

1.4 (incr. in TGP, 250 -> 350 watts) * 1.9 perf per watt = 2.66 (best-case), no?

 

... In selected RT+DLSS scenarios

 

Yeah, best case implies cherry picked.

 

Powering 8K is certainly a very bold claim that should be substantiated by a significant performance increase. The claim of the greatest generational leap in the history of NVIDIA is also a very bold claim that I have never seen NVIDIA make use of in the last 10 years.

 

Can we try to keep the marketing hyperbole to a minimum here during this release?

 

Not really that bold to be honest. Context means everything here.
It could be slightly faster than turning at 8k and it would still make that correct.

 

Bold enough that the same was not mentioned during the Turing unveiling.

 

They could mean 8k DLSS (4k internal resolution in performance mode). No problem for a 3090 even at 60fps.