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]

Not sure where Videocardz got the slide, but keep in mind the 2070 Super is slightly faster than the 5700 XT in Borderlands 3 and Doom Eternal in the quick benchmarks I looked at.

 

[CakeMonster]

The CUDA core count of the 3090 is over 30% better than the 3080, all the other specs except RAM are around 20%. Not sure what to make of that.

 

[JoeRambo]

A100 has 6912 CUDA cores at 1.41GHz to give 19.5 TFlops FP32. It would be pretty incredible if 36 "Shader" TFlops for the 3090 was FP32, but the seems very unlikely.

Depends on the structure of new SM. Remember A100 has to pump FP64 at insane rate, while gaming card does away with that. So 36TF FP32 is actually very possible given the amount of cores they put in and depending on "boost" clocks.

 

[TESKATLIPOKA]

I honestly don't think AMD will be able to compete with 3070 and 3080. You guys are being bit too optimistic considering AMD track record in graphics card. I first need to see a price cut on 5700XT to $300 to start things and $200 on 5600XT.

And you are a bit too pessimistic just because of AMD's past, that doesn't mean It will never change, just look at Bulldozer vs Ryzen.
One thing is competing with RTX 3080, but why do you think Big Navi won't even compete with RTX 3070? RTX 3070 has 2080Ti level of performance or a bit more from what I saw from the marketing slides. Big Navi has 80CU which is double of Navi 10 and RTX 2080Ti was 50% faster than Navi 10. Big Navi should be faster 20-35% faster than RTX 2080Ti, unless 80CU is not true or It has much lower clocks than Navi10. Keep in mind Nvidia wouldn't set the price so low without a reason although the new consoles could be that reason, we will see, I think AMD will need to show us something soon.

 

[TitusTroy]

750w PSU OK with headroom for the 3080?...or is 850w the safer option?...I plan on pairing it with the upcoming Zen 3 CPU (8 core)

 

[IntelUser2000]

Depends on the structure of new SM. Remember A100 has to pump FP64 at insane rate, while gaming card does away with that. So 36TF FP32 is actually very possible given the amount of cores they put in and depending on "boost" clocks.

Then the only explanation is only the shaders scaled, which limits gains. 30-35% gain for 2x Flops is reasonable considering compute isn't the only thing for graphics.

 

[mohit9206]

And you are a bit too pessimistic.
One thing is competing with RTX 3080, but why do you think Big Navi won't even compete with RTX 3070? RTX 3070 has 2080Ti level of performance or a bit more. Big Navi has 80CU which is double of Navi 10 and RTX 2080Ti was 50% faster than Navi 10. Big Navi should be faster 20-35% faster than RTX 2080Ti, unless 80CU is not true or It has much lower clocks than Navi10. Keep in mind Nvidia wouldn't set the price so low without a reason although the new consoles could be that reason, we will see, I think AMD will need to show us something soon.

I mean to say they might not be able to provide better value for money. At same price RTX and Tensor cores give Nvidia an edge and if AMD is cheaper it needs to be significantly cheaper to justify the lack of hardware and software features.

 

[Konan]

Nvidia GeForce RTX 3080 Early Look: Ampere Architecture Performance - Hands-On!

 

[Det0x]

Not sure where Videocardz got the slide, but keep in mind the 2070 Super is slightly faster than the 5700 XT in Borderlands 3 and Doom Eternal in the quick benchmarks I looked at.

So Navi 2x could very well be faster then th 3080 according to this..
2x2070s > 3080

 

[TESKATLIPOKA]

I mean to say they might not be able to provide better value for money. At same price RTX and Tensor cores give Nvidia an edge and if AMD is cheaper it needs to be significantly cheaper to justify the lack of hardware and software features.

AMD will have RTX, I am not sure about a similar tech to Tensor cores and DLSS. They will have to price It accordingly or they will have a problem selling their cards. We will see. I am pretty happy with Ampere except It was not built using 7nm process and the same amount of memory.

 

[Glo.]

No sub-3070 SKUs info, yet

 

[TESKATLIPOKA]

So Navi 2x could very well be faster then th 3080 according to this..
2x2070s > 3080

RTX 2070S is faster than RX 5700 XT and 80CUs doesn't mean 2x more performance than Navi 10 because of not linear scaling. It all depends on clocks and IPC gain If any compared to RDNA1.

 

[lobz]

To be super clear it is a custom 8nm process for NV. Period.

Which means 10nm by nature. Custom or not, SS didn't develop a whole different 8nm node solely for NVIDIA. Just to be super-super clear

 

[JujuFish]

Yep, if it turns out true that the 3070 is a 2080 Ti class GPU for $499 then I'd call that a huge win. Will sell like hotcakes.

The 3070 was a 980 Ti equivalent, but for $379. Of course, the 2080 Ti had a significantly higher MSRP than the 980 Ti, but it just goes to show how skewed people's perspectives are on the 30-series prices. The 20-series being ridiculously expensive doesn't make the 30-series a great deal, or anything approaching "so affordable" IMO.

 

[amenx]

Nvidia GeForce RTX 3080 Early Look: Ampere Architecture Performance - Hands-On!

So the 3080 comes with the 12-pin connector.

 

[Tup3x]

Nvidia GeForce RTX 3080 Early Look: Ampere Architecture Performance - Hands-On!

Oookay... The actual rasterisation performance boost is much higher than I though.

 

[Konan]

I am pretty happy with Ampere except It was not built using 7nm process.

TBH looks like it not being on 7nm is a moot point now. Same for me when Turing was on 12nm. The architecture overall improvements and what looks like perf/$$$ improvements so far seen doesn't really make it a big deal. Looking forward to reviews.

 

[Konan]

So the 3080 comes with the 12-pin connector.

On the Founders Edition yeah.

 

[randomhero]

Videocardz is running pretty hard with the 2X CUDA cores. It will be interesting to see how this all plays out when reviews come out.

NVIDIA announces GeForce RTX 3090, RTX 3080 and RTX 3070 - VideoCardz.com

NVIDIA announces GeForce RTX 30 series NVIDIA today announces its new GeForce RTX 30 series. NVIDIA confirmed that the series based on Ampere architecture are manufactured in Samsung 8nm process technology. The 30 series were announced by NVIDIA CEO Jensen Huang during a special GeForce live...

videocardz.com

We will have to wait for arch deep dive, it seems.

Edit: I still think they can do 2 fp32 ops per clock per alu.

 

[Konan]

Oookay... The actual rasterisation performance boost is much higher than I though.

Looks like it.

Not as big as 1st thought from leaks | Standard Dual Slot card and sightly longer than a FE 2080

Does seem quieter than standard FE 2000 series card

Some of the early comments from that video through the full run of those game benchmarks
Gen Vs Gen 3080 vs 2080 =
80% uplift at 4k in Borderlands demo.
69.8% on Shadow of the Tomb Raider
68%5 on BF5
77.6% on Control
+80% Doom Eternal
92% Quake 2

 

[IntelUser2000]

Nvidia GeForce RTX 3080 Early Look: Ampere Architecture Performance - Hands-On!

This 80% gain corresponds to Shader firepower and memory bandwidth increasing without anything else scaling(such as TMU and ROPs).

A balanced architecture in average will gain roughly a third from each source, 1)Shader throughput 2) TMU/ROP 3) Memory Bandwidth

RTX 3080 has 2.95x the shader throughput and 1.69x the memory bandwidth.

(195% x 0.3) x (69% x 0.3) = 91% gain

Same done to RTX 3080 vs RTX 2080 Ti results in 46%. The RTX 3090 will be a further 20% faster for greater than 70% gain.

 

[Tup3x]

Now that does look sweet...

 

[Stuka87]

TBH looks like it not being on 7nm is a moot point now. Same for me when Turing was on 12nm. The architecture overall improvements and what looks like perf/$$$ improvements so far seen doesn't really make it a big deal. Looking forward to reviews.

Where it does matter is Perf:Watt. Which nVidia pushed on people for years as being extremely important. Its clear now that they are trying to move people away from thinking that. But it doesn't change the fact that having a GPU dump 350-400W of power into your case can cause significant cooling issues for a lot of people.

 

[TESKATLIPOKA]

TBH looks like it not being on 7nm is a moot point now. Same for me when Turing was on 12nm. The architecture overall improvements and what looks like perf/$$$ improvements so far seen doesn't really make it a big deal. Looking forward to reviews.

I wanted better performance/W mainly because I am interested in mobile and not desktop versions.
Performance/W ratio didn't increase much If we consider that a 220W 8GB RTX 3070 has a similar or a bit better performance than a 250W 11GB RTX2080Ti.

 

[MrTeal]

Depends on the structure of new SM. Remember A100 has to pump FP64 at insane rate, while gaming card does away with that. So 36TF FP32 is actually very possible given the amount of cores they put in and depending on "boost" clocks.

¯\_(ツ)_/¯
If that's the case, the card will be a compute monster. 36TFlops FP32 off 28B transistors is a massive increase over the 16.3TFlops from 18.6B transistors that Titan RTX did. It's 50% more FP32 per transistor.