Nvidia Ampere Discussion [2020-05-14]

Do you expect AMD availability to be significantly better post launch?
Who knows what to expect.. it's been a while since AMD has been this much tight lipped about a release.
The RTX 3080 seems similar to the Vega release in volume availability. Big Navi for all we know could either be another Vega or something like a Polaris launch.

If we're to trust @Bondrewd who claims to work for/on a distributor, there'll be plenty of Big Navi boards right after October 28th.
 
Who knows what to expect..
If we expect "Big Navi" to perform similarly to 3080 while having the same (or even lower) price then how exactly would it's availability be better considering that AMD's N7 allocation isn't infinite and the chips themselves will likely be more expensive to produce than GA102?

Big Navi for all we know could either be another Vega or something like a Polaris launch.
To be something like Polaris launch it would have to be sized around 150-200 mm^2 which obviously isn't happening.

If we're to trust @Bondrewd who claims to work for/on a distributor, there'll be plenty of Big Navi boards right after October 28th.
I vaguely remember him saying that there won't be much Navi 21 stock this year at all.

I also kinda wonder about 3080 launch being as Vega in volume - did you see people lining up at stores and estores going down due to load around Vega launch too?
 
If we expect "Big Navi" to perform similarly to 3080 while having the same (or even lower) price then how exactly would it's availability be better considering that AMD's N7 allocation isn't infinite and the chips themselves will likely be more expensive to produce than GA102?
AMD's 7nm allocation on TSMC isn't infinite and obviously neither is nVidia's 8nm allocation on Samsung.
I gather you want to suggest there'll be a lot more GA102 chips than N21 ones, which is a futile discussion because you don't know yields nor waffer allocation on either chip.


To be something like Polaris launch it would have to be sized around 150-200 mm^2 which obviously isn't happening.
There are probably more Polaris chips than N21 chips on the same waffer, yes. Your assumption that the amount of waffers AMD would have for N21 and Polaris is the same or lower is pure speculation on your part.


I also kinda wonder about 3080 launch being as Vega in volume - did you see people lining up at stores and estores going down due to load around Vega launch too?
Yes at least on estores.



I was simply commenting on lack of volume that has been subject of news and memes all over the internet, but it looks like I touched a sensitive subject here which triggered some users and already deviated the conversation about Ampere too much.
I'm dropping this conversation.
 
Last edited by a moderator:
Aren't there reported 7nm supply issues (yield rates issues) at TSMC or is that a rumor?
I think you're referring to the PS5 rumor talking about yield issues on the process, which Sony has outright denied and for which there's no proof or even anything hinting such.
According to TSMC 7nm yields aren't fine, they're better than fine. Their 7nm yields ramped up faster than any process before (5nm is reportedly doing now even better though) and they were at 0.09 defects / cm^2 3 quarters from process's launch (which was around a year ago) https://fuse.wikichip.org/news/2879/tsmc-5-nanometer-update/
 
Seems that the RT cores are not faster Turing's


Do you think it's a code or drivers problem ?


Their analysis makes little sense. The ratio in RT flops to GFLOPS has dropped in the 3080 compared to the 2080, from roughly 2.38 RT flops to GFLOPS (FP32) to roughly 1.95. If the performance drop remains the same, were either looking at improved efficiency per RT flop, or lower utilisation of FP32 GFLOPS, or a mixture of both that cannot be identified with the data presented.
 
Seems that the RT cores are not faster Turing's


Do you think it's a code or drivers problem ?


The infamous Giga rays/second metric mysteriously disappeared, so we don't know how much faster the 3080 should be, based on this metric.
This time the focus was clearly more on traditional rasterization.
 
The infamous Giga rays/second metric mysteriously disappeared, so we don't know how much faster the 3080 should be, based on this metric.
This time the focus was clearly more on traditional rasterization.
They've switched to "RT flops" now, RTX 2080S has "34 RT-flops" while RTX 3080 has "58 RT-flops"
(now we just need to know what on earth those rt flops are and how they compare to MS's "rt equivalent flops" or whatevertheywere, since they're clearly using a different metric)
 
The infamous Giga rays/second metric mysteriously disappeared, so we don't know how much faster the 3080 should be, based on this metric.
This time the focus was clearly more on traditional rasterization.

Path traced minecraft and quake2 rtx are the best performing titles versus turing. Similarly blender can be over 2x faster. It looks like ray tracing is greatly improved in ampere. Probably fps is better metric to compete against amd than gigarays/intersections/...
 
They've switched to "RT flops" now, RTX 2080S has "34 RT-flops" while RTX 3080 has "58 RT-flops"
(now we just need to know what on earth those rt flops are and how they compare to MS's "rt equivalent flops" or whatevertheywere, since they're clearly using a different metric)
I think they've actually said that they are using MS metric for this?
 
Path traced minecraft and quake2 rtx are the best performing titles versus turing. Similarly blender can be over 2x faster. It looks like ray tracing is greatly improved in ampere. Probably fps is better metric to compete against amd than gigarays/intersections/...

These games are not very representative for normal today’s games as they have very simple geometry which fits nicely in the GPU caches.
For the 3080 L1 caches have grown, but L2 cache was reduced relative to the 2080Ti.
 
Path traced minecraft and quake2 rtx are the best performing titles versus turing. Similarly blender can be over 2x faster. It looks like ray tracing is greatly improved in ampere. Probably fps is better metric to compete against amd than gigarays/intersections/...

I think the hardware reviewers sometimes forget that they're running software benchmarks and not hardware benchmarks. They're really testing how well the software runs and then trying to extrapolate that to hardware performance by looking for trends. It should have been fairly obvious that if they wanted to target the performance of the RT cores, then running benchmarks in games is probably not the best way, but if they were going to test with games then (as you said) fully path-traced games like Quake 2 and Minecraft may have been a smarter option. Ultimately, when you ray-trace the amount of shading you do per-pixel goes way up and you waste cache performance. Doubling the speed of the RT cores would likely not significantly close the gap between raster and ray-traced performance, because it is probably not the bottleneck.
 
These games are not very representative for normal today’s games as they have very simple geometry which fits nicely in the GPU caches.
For the 3080 L1 caches have grown, but L2 cache was reduced relative to the 2080Ti.

If you want to measure rtcore performance you need rtcore heavy games. Other games might not use enough ray tracing for rtcore to become so serious bottle neck that the difference is blatantly obvious.

Another thing is that ampere can run rtcore better in parallel to other loads. Utilizing better parallel compute+rtcore could require patches to games. i.e. difference over turing might become larger once games are ampere optimized. It will be super interesting to see how cyberpunk2077 performs.
 
This isn't a case of demand > volume. It's a $700 card and there aren't that many people who are capable/willing to spending that much money on a GPU for games.

You recall up to 100 people lining up outside of stores for a $700 GPU lunch, like it's a new iPhone? Cause I sure don't.


https://www.mooreslawisdead.com/post/nvidia-s-ultimate-play

I will start this post by getting straight to the point – I have evidence that suggests Nvidia is trying to have their cake and eat it too when it comes to the perceived price/performance of their new Ampere RTX 30-Series lineup. They are attempting to appear to be launching a lineup that is priced lower than their much maligned Turing generation, but in reality these things will cost far more than they are letting on for the overwhelming majority of shoppers this fall.

I am sorry, but that guy has proven himself misinformed at best, and stupid at worst, so many times now. How does a no-name youtuber mumbling nonsense into the camera rise to any sort of prominence anyway?

We would be better served simply looking at Nvidia's financial forecast: +$530M Q/Q , (+$430M consensus), +1.4B Y/Y which, even accounting for Mellanox is still a substantial increase. Especially considering that various Turning SKUs have been EOLed, this very strongly suggests substantial Ampere shipments this quarter.

edit: spelling
 
Last edited:
If you want to measure rtcore performance you need rtcore heavy games. Other games might not use enough ray tracing for rtcore to become so serious bottle neck that the difference is blatantly obvious.

Another thing is that ampere can run rtcore better in parallel to other loads. Utilizing better parallel compute+rtcore could require patches to games. i.e. difference over turing might become larger once games are ampere optimized. It will be super interesting to see how cyberpunk2077 performs.

Synthetic benchmarks designed specifically to bottleneck the RT cores would probably be the best way. Not sure if those exist right now.
 
Back
Top