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.

 

[jpiniero]

I wonder if they used DX11 or DX12 in the Borderlands 3 test.

The 2070 and 1070 result seems to match up more with DX11. Also note NV used an unnamed i9 processor, it could be just about anything including a Rocket Lake ES.

 

[insertcarehere]

https://twitter.com/x/status/1302023768613761024

RTX 3070 even according to Nvidia materials in Unreal Engine 4 game is 43% faster than RX 5700 XT in 1440p.

Thats not really painting good picture for Ampere, does it?

Nvidia's graphs only compared 3070 to 2070/1070 in Borderlands 3 (they didn't do RDNA comparisons at all), the numbers that this poster is quoting is only done by comparing Nvidia's benchmark numbers to Techspot's even though the tests are obviously not done in identical conditions (Techspot for one didn't use the in-game benchmark tool while Nvidia most likely would). Let's wait for actual reviews to come out before claiming one way or another.

 

[Kenmitch]

But nVidia has covered every corner.

Top speed is always achieved in the straight away!

 

[Gideon]

Nvidia's graphs only compared 3070 to 2070/1070 in Borderlands 3 (they didn't do RDNA comparisons at all), the numbers that this poster is quoting is only done by comparing Nvidia's benchmark numbers to Techspot's even though the tests are obviously not done in identical conditions (Techspot for one didn't use the in-game benchmark tool while Nvidia most likely would). Let's wait for actual reviews to come out before claiming one way or another.

Well it is a speculation thread

But yeah, overall I agree. The results vary enough between reviewers that you can't really read too much into it. 2080 Ti for instance got 90FPS @ 1440p according to techspot, but only 78.5 FPS according to techpowerup. That delta is so wide that it 3070 can be both faster or slower than the 2080 Ti

 

[Konan]

Because Nvidia needs something to compete with Navi 22.

And reusing 102 dies are too expensive.

If RTX 3080 is 68 SMs with 320 Bit GDDR6X, then 103 with 60 SMs and 256 bit GDDR6X will get you around 85% of RTX 3080 performance.

And don't forget. Already with 2944 ALUs/5888 CUDA cores RTX 3070 is 220W TDP, and full 48 SM with GDDR6X would be 250W TDP.

Don't expect 60 SM/256/320 Bit design to be any less than 300W on 8 nm process.

And that is not going to compete with 60 CU RDNA2 GPU that will use around 200-225W of power.

Nvidia has gone first and can now cut and segment as much as they want. There is tons of room. Between the 3070 and 3080 there is up to 30-40% in space.
Please don't forget that the 3070 has DLSS and RT too! Power doesn't matter when you have performance advantage. Remember the 3070 minimum will be +7% above a 2080Ti.

PS. Enjoy your time off?

 

[Konan]

NVIDIA_DEV.2204 = "NVIDIA GeForce RTX 3090"
NVIDIA_DEV.2206 = "NVIDIA GeForce RTX 3080"
NVIDIA_DEV.222B = "NVIDIA GeForce RTX 3080 Ti Engineering Sample"
NVIDIA_DEV.222F = "NVIDIA GeForce RTX 3080 11GB Engineering Sample"
NVIDIA_DEV.223F = "NVIDIA GA102GL"
NVIDIA_DEV.2482 = "NVIDIA GeForce RTX 3070 SUPER"
NVIDIA_DEV.2484 = "NVIDIA GeForce RTX 3070"
NVIDIA_DEV.24AF = "NVIDIA GeForce RTX 3070 Engineering Sample"
NVIDIA_DEV.24BF = "NVIDIA GeForce RTX 3070 Engineering Sample"
NVIDIA_DEV .252F = "NVIDIA GeForce RTX 3060 Engineering Sample"

Pulled from the current PCI ID Repository showing added support.

 

[Glo.]

Nvidia has gone first and can now cut and segment as much as they want. There is tons of room. Between the 3070 and 3080 there is up to 30-40% in space.
Please don't forget that the 3070 has DLSS and RT too! Power doesn't matter when you have performance advantage.

Let me repeat it one more time.

The fact that Nvidia develops 60 SM GPU is because its too expensive to cut 102 dies down to Navi 22 levels of performance, and 104 die is way too weak to compete with this GPU. EVEN WITH DLSS.

So where do you get that idea of performance advantage?

 

[sontin]

Navi22 is 3070 level. You need to stop with these baseless claims. Navi with 60CUs makes it barely to the 3070. 3070 will be ~60% faster than the 5700XT with only 46/48SM. GA104 is enough for everything AMD will throw at them.

 

[Konan]

Let me repeat it one more time.

The fact that Nvidia develops 60 SM GPU is because its too expensive to cut 102 dies down to Navi 22 levels of performance, and 104 die is way too weak to compete with this GPU. EVEN WITH DLSS.

So where do you get that idea of performance advantage?

Yes I saw what Kopite tweeted. For Nvidia is there is still some cost advantage from cutting 102 as an option because of the Samsung node. Don't forget GA103 for the future.
The card that AMD drops in priced closed to the 3070 will be the card it competes with. Nvidia will simply cut from wherever to compete with the next round.

 

[Kenmitch]

Geekbench Results

ASUS System Product Name - Geekbench

Benchmark results for an ASUS System Product Name with an Intel Core i9-10900K processor.

browser.geekbench.com

table source

Erster Geekbench-Wert zur GeForce RTX 3080 aufgetaucht | 3DCenter.org

In unserem Forum wurde der erste Eintrag einer GeForce RTX 3080 Grafikkarte unter dem Geekbench 5 bzw. dem dortigen OpenCL-Test erspäht (Update: CUDA-Werte hinzugefügt). Wie bei den bereits vorliegenden Benchmark-Werten einer GeForce RTX 3080 unter dem

www.3dcenter.org

	Hardware	GB5 OpenCL	Anmerkung
GeForce RTX 3080	Ampere, GA102, 68 SM @ 320 Bit GDDR6X	191'891	+86% vs. 2080, +74% vs. 2080S, +46% vs. 2080Ti
GRID V100S-32Q	Volta, GV100, 80 SM @ 4096 Bit HBM2	153'272	-
Titan RTX	Turing, TU102, 72 SM @ 384 Bit GDDR6	134'369	-
GeForce RTX 2080 Ti	Turing, TU102, 68 SM @ 352 Bit GDDR6	131'494	+28% vs. 2080, +19% vs. 2080S
GeForce RTX 2080 Super	Turing, TU104, 48 SM @ 256 Bit GDDR6	110'399	+7% vs. 2080
GeForce RTX 2080	Turing, TU104, 46 SM @ 256 Bit GDDR6	102'969	-

We don't even know if the current drivers are optimized for Geekbench yet....Wait for the official reviews and launch drivers is best!

 

[Asterox]

Hm, 30% higher score with allegedly 28% higher power consumption.

https://twitter.com/x/status/1303014612397436935

 

[jpiniero]

Hm, 30% higher score with allegedly 28% higher power consumption.

https://twitter.com/x/status/1303014612397436935

If you look at the tests themselves, 2 of them are barely faster and 1 is actually slower. 5 of the tests are ~50% faster. The other 3 are around 30%.

 

[Konan]

Only 1 test/just 1 run leaks - Ashes of the Singularity score - for 3080 - puts it 27% above a bunch of average runs of 2080ti. (prob beta drivers too)
Now we know that reviews have the cards

NVIDIA GeForce RTX 3080 4K AotS benchmark leaks out - VideoCardz.com

NVIDIA GeForce RTX 3080 turning ashes into frames A leak from the Ashes of the Singularity database featuring GeForce RTX 3080 graphics cards has been discovered by @_rogame. The graphics card is possibly currently being tested by a reviewer. The GeForce RTX 3080 graphics card features 8704 CUDA...

videocardz.com

 

[Zstream]

Only 1 test/just 1 run leaks - Ashes of the Singularity score - for 3080 - puts it 27% above a bunch of average runs of 2080ti. (prob beta drivers too)
Now we know that reviews have the cards

NVIDIA GeForce RTX 3080 4K AotS benchmark leaks out - VideoCardz.com

NVIDIA GeForce RTX 3080 turning ashes into frames A leak from the Ashes of the Singularity database featuring GeForce RTX 3080 graphics cards has been discovered by @_rogame. The graphics card is possibly currently being tested by a reviewer. The GeForce RTX 3080 graphics card features 8704 CUDA...

videocardz.com

View attachment 29381

About right IMO.

 

[Krteq]

Hm, 30% higher score with allegedly 28% higher power consumption.

Spoiler

https://twitter.com/x/status/1303014612397436935

So, cut-down GA102 GPU is about 30% faster than cut-down TU102 with approx. 30% higher power consumption?

Another interesting bits

https://www.3dcenter.org/news/erster-geekbench-wert-zur-geforce-rtx-3080-aufgetaucht

 

[Tarkin77]

+0% Performance per Watt

 

[Krteq]

Seems so

but...but.... "1.9x perf/watt" (normalized to 60 FPS in Control)

 

[badb0y]

HMMMM. Let's see what the benchmarks say on the 14th...

 

[Zstream]

+0% Performance per Watt

What did you expect? The node process and nothing really new to add in terms of hardware leaves little to the imagination. It’s a good raw performance increase, so not sure what people want.

If and when something new comes along in terms of hardware, let’s talk. Otherwise it will be the the race to use RT in everything.
I still wish physx was able to be offloaded like RT is.

 

[Konan]

3080 20GB & 3070 16GB SUPER’s incoming ? October 24??

https://twitter.com/x/status/1303044982459097089

 

[xpea]

Another interesting bits

https://www.3dcenter.org/news/erster-geekbench-wert-zur-geforce-rtx-3080-aufgetaucht

All I see is a looking forward architecture that is a monster in compute workloads like RT / ML and that will age very well.
Once Nvidia will help devs to optimize their engined for Ampere (ie less INT instructions, more FP32), the difference will be much more pronounced than now.
30 FP32 TFOPS for $700 in 2020? Wow wow wow

 

[itsmydamnation]

All I see is a looking forward architecture that is a monster in compute workloads like RT / ML and that will age very well.
Once Nvidia will help devs to optimize their engined for Ampere (ie less INT instructions, more FP32), the difference will be much more pronounced than now.
30 FP32 TFOPS for $700 in 2020? Wow wow wow

are you serious?
you always post things like this do you ever have an actual clue about what you are saying?

So you want game dev's to take INT operations and run them as FP operations at reduced range and resolution and at increased power for 0 performance improvement?

 

[xpea]

are you serious?
you always post things like this do you ever have an actual clue about what you are saying?

So you want game dev's to take INT operations and run them as FP operations at reduced range and resolution and at increased power for 0 performance improvement?

Oh so ratio int/fp is all the same between games?

 

[Bouowmx]

3080 20GB & 3070 16GB SUPER’s incoming ? October 24??

https://twitter.com/x/status/1303044982459097089

So NVIDIA is launching like the entire lineup in 2-3 months? RTX 3050 family in November?
Kind of hinting to Ampere GeForce being a short generation.

 

[maddie]

All I see is a looking forward architecture that is a monster in compute workloads like RT / ML and that will age very well.
Once Nvidia will help devs to optimize their engined for Ampere (ie less INT instructions, more FP32), the difference will be much more pronounced than now.
30 FP32 TFOPS for $700 in 2020? Wow wow wow

Oh so ratio int/fp is all the same between games?

Serious posts?
Do you understand what an (either/or) (int/fp shader) can do? Why would you think using less int ops is so great? Its only if int ops are greater than 50% of all ops will you get idle resources.