Nvidia Ampere Discussion [2020-05-14]

you think the 320W is for a 20 GB model? Otherwise 3080 is 10 GB and 3070 is 8 GB, which shouldn’t be that much of a difference.

 

I don’t understand the relevance other than number of SMs, RT and tensor cores, or if they’re built on different nodes. The fact that 3070 has the same number of cores as a 2080 super and 3080 has the same number of cores as a 2080 ti confuses me because the power consumption difference has grown to 100W. I can only understand it if the 3080 has a greater ratio of RT and tensor cores, or if it’s built on a much worse node.

 

No 10GB GDDR6X vs 8GB GDDR6. Huge difference there.

 

I didn’t notice that difference. I thought they were all gddr6x

 

I can sort of see the 3080 to 3070 power draw difference. On memory bandwidth alone, you have 25% more from extra channels which also use GDDR6X vs 6, which is 48% more data being moved around. Add in more SMs running at a higher clock.....

50% of 220W = 110W + 220W = 330W. So given the alleged specs, the power difference seems reasonable. But the 3080 providing 50% more performance over a 3070 seems like an unusually large gap.


EDIT:

If one looks at the chip tier list, it makes a bit more sense:

Titan -> 3090 (very large chip)
Ti -> 3080 (very large chip)
*80 -> *70 (large chip)

 

@ninelven yah looking at it the 3070 will probably have roughly 2/3 the number of cores. So the numbers start to line up. 2070 to 2080 (edit: wrote 3080 here) was about 4/5.

 

One of the twitter leaker wrote that the 3090 (82SMs, 24GB, 920GB/s) has 350W and the 3070(TI) with 3072 cores has 250W. That would put the 3090 in another efficiency league than the 3070(TI)...
For comparision: RTX2070FE had a 175W TDP and the RTX2080TI FE 270W. The difference was bigger...

 

Looking good. And only 2 slots too.

 

Yeah the angle threw me off. It's probably pushing 3 slots like most big AIB cards.

It'll be interesting to see how the founders edition compares in temps and clocks. Nvidia is jumping through all sorts of engineering hoops on the FE version but so far AIB models look like standard fare.

 

Picture with a GA102 chip:

 

A 3070 ti would make a lot of sense, because there's a perfect spot at 270W to fit another card into the lineup.

 

Return of pre-Fermi separated clock domains or fake?

 

Those look like dummy numbers for show only. 1234, 4321, 2345....

 

lol... I didn't noticed, thx

 

Nvidia's old strategy: We're so dominant we can claim low power usage as something to care about!
Nvidia's new strategy:

 

It only becomes an issue if AMD comes up with something as fast or faster and more power efficient. Especially for mobile.

 

I'm thinking this is very possible for non-DXR/RTX games. Transistor density has increased much more than the SM count has. I have a feeling a lot of that is going to RT cores and tensor cores, but hopefully a good chunk to making the SMs wider, or something that can improve general compute performance. I have a feeling ampere will show great improvements in RT performance, especially with DLSS, but performance in non-DXR/RTX games will not improve as greatly. In a sense, like RTX 20x0 series, I think they're betting on the future. Curious to see how it plays out with UE5 and Cryengine, which seem to be doing things like GI without using DXR on PC.