Question 'Ampere'/Next-gen gaming uarch speculation thread

[Ottonomous]

How much is the Samsung 7nm EUV process expected to provide in terms of gains?
How will the RTX components be scaled/developed?
Any major architectural enhancements expected?
Will VRAM be bumped to 16/12/12 for the top three?
Will there be further fragmentation in the lineup? (Keeping turing at cheaper prices, while offering 'beefed up RTX' options at the top?)
Will the top card be capable of >4K60, at least 90?
Would Nvidia ever consider an HBM implementation in the gaming lineup?
Will Nvidia introduce new proprietary technologies again?

Sorry if imprudent/uncalled for, just interested in the forum member's thoughts.

 

[AtenRa]

That is still wrong. TSMC's 7nm is about density and not clock rate. Look at every SoC produced on it. All companies have been gone wide with a low clock rate.

And 1,7GHz with 34,5B xtors is very high. That is Turing level with an density increase of 2,2x.

It is obvious that they made a decision to go on high density vs speed much like A100, that is why we see 350W TDP at only 1.7GHz

Yea you are getting Turing speed at high density but with 40% higher TDP (350W vs 250W).

 

[Kenmitch]

3070 no gddr6x ???? shite 8nm, for first time in past 20 years AMD here I come!

NV root in hell

It's probably best to just wait it out for the official launch and reviews from both camps.

I doubt I'm buying either offering, but at least the reviews will be interesting....Of course the real actions in the comment section.

 

[Glo.]

It may still be possible, that its 7 nm EUV from Samsung, and not TSMC.

For the density part.

 

[sontin]

It is obvious that they made a decision to go on high density vs speed much like A100, that is why we see 350W TDP at only 1.7GHz

Yea you are getting Turing speed at high density but with 40% higher TDP (350W vs 250W).

No, you are getting twice the transistors and 50% more bandwidth with 40% higher TDP.

 

[CastleBravo]

We will find out next week. Either NV will give us all the details, or they will beat around the bush and we will know their hand is weak.

 

[AtenRa]

No, you are getting twice the transistors and 50% more bandwidth with 40% higher TDP.

GTX980ti at 28nm to RTX2080Ti 12nm

We got more than double the transistors
We got double the Memory Bandwidth
+ we got 50% increased boost clocks

At the same 250W TDP

 

[DDH]

No, you are getting twice the transistors and 50% more bandwidth with 40% higher TDP.

On a smaller more efficient node, yet it has an increase tgp of 40%.

However the a100 has 54B transistors, 8192 cuda cores, with a 90% tdp of 250w or max of 400w. But a die rumoured to have 50% less transistors, supposedly on the same node will have a tgp almost as high

Sent from my SM-N975F using Tapatalk

 

[Glo.]

No, you are getting twice the transistors and 50% more bandwidth with 40% higher TDP.

50% more bandwidth you WOULD get if you would go from 14 Gbps to 21 Gbps.

And according to info VCZ got its only 19.5 Gbps. Which means 40% more bandwidth.

 

[sontin]

GTX980ti at 28nm to RTX2080Ti 12nm

We got more than double the transistors
We got double the Memory Bandwidth
+ we got 50% increased boost clocks

At the same 250W TDP

Nope. Its 270W against 225W. GTX980TI was thermal limited by the radial cooler: https://www.hardware.fr/articles/935-4/consommation-efficacite-energetique.html

 

[Kenmitch]

50% more bandwidth you WOULD get if you would go from 14 Gbps to 21 Gbps.

And according to info VCZ got its only 19.5 Gbps. Which means 40% more bandwidth.

You're doubting the word of Sontin? He lives, breathes, and sleeps all things Nvidia. If anybody knows the facts it's him.

 

[Mopetar]

I'm guessing NV knows that anyone willing and able to buy a 3090 would never put up with loud fans. So what if they worked out how to make this funky cooler silent?

NVidia users are used to really loud fans. They're always going on about how great the company's cards are. Normally they're pretty cool , but sometimes they do blow a lot of hot air though.

Better than the AMD fans though. They might not always be as loud, but the last few years they've been constantly whining about how things will be better next time.

 

[AtenRa]

Nope. Its 270W against 225W. GTX980TI was thermal limited by the radial cooler: https://www.hardware.fr/articles/935-4/consommation-efficacite-energetique.html

Even so, we got way more from 980Ti to RTX2080Ti with only 20% more power consumption than we are getting from 2080Ti to RTX3090 with 40% increase of power.

Not only that, but if RTX3090 is 60% faster vs RTX2080Ti in Raster performance but with a 40% increase of power, then they only managed a 15% increase of perf/watt from Turing to Ampere ????

with a full node ???? wtf

 

[n0x1ous]

Even so, we got way more from 980Ti to RTX2080Ti with only 20% more power consumption than we are getting from 2080Ti to RTX3090 with 40% increase of power.

Not only that, but if RTX3090 is 60% faster vs RTX2080Ti in Raster performance but with a 40% increase of power, then they only managed a 15% increase of perf/watt from Turing to Ampere ????

with a full node ???? wtf

980Ti to 2080Ti is 2 full nodes. 2080ti to 3090 is 1 full node. comparison makes no sense

 

[sontin]

There doesnt exist any performance numbers outside of Timespy Extreme. So just wait for the reveal. Without performance and real world power numbers it just useless to discus. And btw: My 2080TI with 360W hits only 7400 points. So a 3090 would be around 40% more efficient...

 

[AtenRa]

980Ti to 2080Ti is 2 full nodes. 2080ti to 3090 is 1 full node. comparison makes no sense

Nope, its only a single node, 12nm is almost the same as 16nm

 

[Stuka87]

That is still wrong. TSMC's 7nm is about density and not clock rate. Look at every SoC produced on it. All companies have been gone wide with a low clock rate.

Not sure where you are getting this. Ryzen isn't clocked low. PS5 and XBSX aren't clocked low. Phones based on it aren't clocked low.

 

[Glo.]

Not sure where you are getting this. Ryzen isn't clocked low. PS5 and XBSX aren't clocked low. Phones based on it aren't clocked low.

Renoir has all three. High density, 60 mln xTors/mm2, efficiency and high clock speeds.

 

[sontin]

SoC GPUs are clocked very low on TSMCs process. Look at Kirin980 to Kirin990 or Qualcomm's transition from Samsung's 10nm LPP to TSMC 7 process. They all went wide and reduced the clock rate.

And XBSX has around ~12b xtors. That is 1/3 of GA102.

 

[tviceman]

If the specs are accurate it should achieve theses scores with 9% faster Clocks, 20% more shaders and 50% more bandwidth. This could only be possible with some major architectural improvements.

EDIT
For comparison 2080Ti is only 56% faster than 2070 in 4K, even though it has 80% more shaders and 37% more memory bandwidth, though yeah, it has slightly slower clock at -5%.

Architectural deficiencies? How does the 1080 TI compare vs. the 1070 or 1080? Given if there are similar resource differences to your 2080 TI and 2070 comparison, I'll be the performance Delta's are equally similar.

I think it's more valid to compare generational leaps (i.e. Pascal vs. Maxwell) than comparing within an existing family (Turing vs. Turing).

 

[AtenRa]

There doesnt exist any performance numbers outside of Timespy Extreme. So just wait for the reveal. Without performance and real world power numbers it just useless to discus. And btw: My 2080TI with 360W hits only 7400 points. So a 3090 would be around 40% more efficient...

OC vs Default ?? come on you can do better than that, I mean its like I would say that RTX2080Ti undervolt has higher perf/watt that Ampere 3090

 

[Stuka87]

SoC GPUs are clocked very low on TSMCs process. Look at Kirin980 to Kirin990 or Qualcomm's transition from Samsung's 10nm LPP to TSMC 7 process. They all went wide and reduced the clock rate.

And XBSX has around ~12b xtors. That is 1/3 of GA102.

A GPU on it's own isnt an SoC....

But doesn't change the fact that other GPU's have high clock speeds, as in, 2+GHz.

 

[Glo.]

SoC GPUs are clocked very low on TSMCs process. Look at Kirin980 to Kirin990 or Qualcomm's transition from Samsung's 10nm LPP to TSMC 7 process. They all went wide and reduced the clock rate.

And XBSX has around ~12b xtors. That is 1/3 of GA102.

XDDDDD

2.1 GHz VEGA iGPU from that hery Renoir chip says hi to your claims

Power of engineering. Making 2.1 GHz iGPU fast AND efficient on 7 nm node, with 60 mln xTors/mm2 density.

 

[JasonLD]

https://twitter.com/x/status/1299313471297814528

https://twitter.com/x/status/1278503357804457984

Kopite was such a bad leaker that he nailed every single thing he leaked.

Well?

 

[jpiniero]

https://twitter.com/x/status/1299313471297814528

https://twitter.com/x/status/1278503357804457984

Wouldn't be wrong if nVidia just calls SS8 "7 nm"

 

[Glo.]

https://twitter.com/x/status/1299313471297814528

https://twitter.com/x/status/1278503357804457984

Well?

I think Kopite is right here, as well .