Nvidia Ampere Discussion [2020-05-14]

Both 3090 and 3080 are based on the GA102.
The former with 7 GPCs / 112 ROPs the latter with 6 GPCs / 96 ROPs.
We need to have that white paper (or reviews) to be sure.

 

Geez what kinda drugs do you need to be on to play that game. I almost had a seizure just watching the video.

 

The only drug you need is Doom Eternal.

 

Was hoping for possibly VP9 encode acceleration from NVENC.

The 32x FP32 or 16 x FP32 INT32 arrangement was what many were speculating when first rumored so it's not surprising. Although it's interesting that they went for still separate FP32 and INT32 setup as opposed to a FP32/INT32 setup.

The performance uplift also has some interesting implications from architecture to architecture. With Pascal->Turing in theory the higher percentage of INT32 relative to FP32 (well upto 50%) in your workload the higher the performance uplift. But with Turing->Ampere in theory it's the reverse, the lower the percentage of INT32 relative to FP32 the more uplift.

 

Came accross this video. No idea if this guy has any reputation but there is it:

Interesting part starts at 11' with RTX30 comparison to big Navi

 

Sounds like rolling up the most negative rumors regarding RDNA2 floating around currently and then adding some, wouldn't put much weight on it.

 

It's the same guy insisting on a RT co-processor. Completely clueless guy.

 

This type of rumors (co-processors, EDRAM, etc.) have been a recurring phenomena ever since G80. It's an easy click-bait wishful fantasy.

 

310w at 100%... ouch. There's probably going to be a bunch of undervolting by users.

 

Thanks for sharing.

I think that power limit is wise to not mess with. We have to remember that the Samsung 8nm node have never been used for such a large die, so nVidia could well be pushing it beyond what is was designed/intended for in the first place. They are probably happy that they can even stay at 310 watts, LOL

 

If I get a 3080 I'm going to be under-volting. To keep up with that heat my case fans will have to work pretty hard. I'd rather keep my fans as low as possible than overclock to get an extra 10% performance.

 

This is why I like AIOs on my GPUs. Mount the rad in the back or top of the case and send the heat straight out. Hopefully someone will release a 240mm hybrid card soon so I don't have to a custom one.

The market is saturated with CPU AIOs optimizing for the corner case (synthetic benchmarks) and yet GPUs run at 100% all day every day stuck on air coolers that give off 300W of heat into your case.

 

Yah, the GPU is by far the most heat, so it makes sense to try to get all of that heat out immediately and not have it affect the other components. CPU heat is negligible by comparison and if some of the cpu heat flows through the rad it's not as big a deal.

 

I'm using a Kraken G12 AIO adapter on my 1080Ti with a 240mm rad mounted on the top of the case and it keeps it under 60C easily even at 110% power limit with a custom overclock curve. But my 5820K @ 4.4 GHz is doing just fine with an ancient TRUE 120 air cooler.

 

Igor from Igor's Lab released a video stating that there will be a shortage of Founder's Editions of the 3080/3090 at launch. But it is not due to insufficient wafers, but due to the complicated coolers. Although he was told Nvidia should be catching up fast and by October, FE's and AIB models should be available in large quantities.

 

Huh? So nVidia is selling "automatic overclocking"? Later in the video the 3080 runs at 1680Mhz with 280W...

 

On desktop though? Looks like it's running the scan in GFE. It tests different points on the voltage frequency curve so working as it should I believe.

Btw, review dates:

 

https://www.techpowerup.com/review/...ture-board-design-gaming-tech-software/4.html
https://hothardware.com/reviews/nvidia-geforce-rtx-30-series-ampere-details
Architecture articles started to pop up
It seems the "minimal" RT core architecture is what we will see in NAVI.

 

From the hothardware link:

"With Ampere, NVIDIA wanted to be able to process Bounding Box and Triangle intersection rates in parallel. So, Ampere’s separate Bounding Box and Triangle resources can run in parallel, and as mentioned, Triangle Intersection rates are twice as fast.

A new Triangle Position Interpolation unit has also been added to Ampere to help create more accurate motion blur effects."

Sounds like pretty significant speedups to ray tracing. The triangle position interpolation is interesting. So frame to frame they can interpolate a new position and if it's outside the current triangle intersection then I guess they can go and test a new intersection. Different from having to test the bounding box and triangle intersection every frame.