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

Page 81 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

Ottonomous

Senior member
May 15, 2014
559
292
136
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.
 

Stuka87

Diamond Member
Dec 10, 2010
6,240
2,559
136
I REALLY doubt we will see all those Ti models at launch. If anything, those will come out as Super versions 6 months down the road.

Also doubt the 192 ROPs count. 2080Ti has 88, going to 192 seems like an utter waste. There is no way going to so many would offer any kind of benefit.
 

IntelUser2000

Elite Member
Oct 14, 2003
8,686
3,786
136
Rumor / Speculation...
We'll see how close they come ;)

Equivalent performance is equivalent to Turing architecture 2304 level PS: basically it is 2070 but it is very cheap. Price is ~350-375 US dollars

LOL, $350 US is cheap?!?

Equivalent performance is equivalent to Turing architecture 3456 level PS: this is OK GA104 cut version. Better than 2080 Super. Within 17%

What's with quoting weirdly specific numbers? Why not add another decimal?

20-30% increase in perf/$ if that is correct. That points to yet another price increase but keeping names same to have gamers accept the new reality.
 
  • Like
Reactions: psolord

Bouowmx

Golden Member
Nov 13, 2016
1,147
551
146
2000 yuan (290 USD) for the lowest 3060. And under 3000 yuan (guess: 2500 yuan/360 USD?) for the next higher 3060. Nice: RIP me.
Disclaimer: I have no idea how China hardware prices work (ex. if they have VAT included like in Europe), or if they can be directly converted to USD.
 
  • Like
Reactions: Konan

Konan

Senior member
Jul 28, 2017
360
291
106
What's with quoting weirdly specific numbers? Why not add another decimal?

20-30% increase in perf/$ if that is correct. That points to yet another price increase but keeping names same to have gamers accept the new reality.

I just copy and pasted what was posted From a translation.
 

IntelUser2000

Elite Member
Oct 14, 2003
8,686
3,786
136
I just copy and pasted what was posted From a translation.

I know, I'm not blaming you. It's like Intel saying 18% for Sunny Cove and AMD saying 52% for Zen. Really? Did everybody forget ranges exist and can be used?

Looks like Nvidia hasn't learned from Turing if the prices are true, but who knows?

The gap between RTX 3080 and 3090 seems way too large. The RTX 3080 performs 30% better than RTX 2080 Ti, but the RTX 3090 that's barely 10% better than RTX 3080 is $1999? So the RTX 3080 is $1299 for a zero perf/$ gain? Or are they going to insert a RTX 3080 Ti that's 5% faster but costs $250 more?
 
  • Like
Reactions: psolord and Konan

reb0rn

Senior member
Dec 31, 2009
240
73
101
With those prices NV can take those cards and push them where sun do not shine.....
also TDP is insane...... just no tnx, good bye PC gaming let it die already
 

GodisanAtheist

Diamond Member
Nov 16, 2006
7,166
7,666
136
I know, I'm not blaming you. It's like Intel saying 18% for Sunny Cove and AMD saying 52% for Zen. Really? Did everybody forget ranges exist and can be used?

Looks like Nvidia hasn't learned from Turing if the prices are true, but who knows?

The gap between RTX 3080 and 3090 seems way too large. The RTX 3080 performs 30% better than RTX 2080 Ti, but the RTX 3090 that's barely 10% better than RTX 3080 is $1999? So the RTX 3080 is $1299 for a zero perf/$ gain? Or are they going to insert a RTX 3080 Ti that's 5% faster but costs $250 more?

- I think it's safe to say that leaker is just some guy spitballin based on updated info.

That ROP count is absolutely outrageous unless NV is redefining what a ROP is or they plan on pushing 16k displays or something (VR maybe?)
 
  • Like
Reactions: ozzy702 and FaaR

FaaR

Golden Member
Dec 28, 2007
1,056
412
136
That ROP count is absolutely outrageous unless NV is redefining what a ROP is or they plan on pushing 16k displays or something (VR maybe?)
They're gonna need way more than a mere 1TB/s memory BW to effectively use 192 ROPs... Sounds like a fanboy making up his personal wet dream numbers to me. ;)
 
  • Like
Reactions: Gideon

Glo.

Diamond Member
Apr 25, 2015
5,803
4,777
136

Mopetar

Diamond Member
Jan 31, 2011
8,113
6,768
136
Drivers would seem to be the most obvious culprit assuming it's a legitimate entry to begin with.

I'm not sure if I'd put much stock in this given how out of line with any reasonable expectations it is.
 
  • Like
Reactions: NTMBK

uzzi38

Platinum Member
Oct 16, 2019
2,705
6,427
146
Drivers would seem to be the most obvious culprit assuming it's a legitimate entry to begin with.

I'm not sure if I'd put much stock in this given how out of line with any reasonable expectations it is.
It's almost definitely legitimate. Just as you said, just driver weirdness. The VideoCardz article explains why they all think its the 3080 very well.
 
  • Like
Reactions: Glo.

Konan

Senior member
Jul 28, 2017
360
291
106
HYPE TRAIN!

And they were saying AMD rumors are complete BS, here came the Nvidia fanboys on twitter/baidu...

Maybe.....
Here is a performance chart from Chiphell. Completely difference source / person / area..... corroborating some of the prior stuff I posted last page.
It is my feeling that the 3070Ti will be trading with Big Navi.

The graph is vague, yes. But it is what it is and what we have for now. However, I think it isn't true as it is too early for it to be real.



111.png
 
Last edited:
  • Love
Reactions: psolord

uzzi38

Platinum Member
Oct 16, 2019
2,705
6,427
146
Just like the 1080Ti 10GB...

Man, I never knew that thing died the day before launch, I thought Nvidia upped the spec long that. Kinda funny how TPU still has it in it's database.
 

Konan

Senior member
Jul 28, 2017
360
291
106
Just like the 1080Ti 10GB...

Man, I never knew that thing died the day before launch, I thought Nvidia upped the spec long that. Kinda funny how TPU still has it in it's database.

Wow I forgot about that. Crazy last minute changes.

The whole 10GB Vs 20GB on the 3080 is going to play out now.
 

TESKATLIPOKA

Platinum Member
May 1, 2020
2,523
3,037
136
Maybe.....
Here is a performance chart from Chiphell. Completely difference source / person / area..... corroborating some of the prior stuff I posted last page.
It is my feeling that the 3070Ti will be trading with Big Navi.

The graph is vague, yes. But it is what it is and what we have for now. However, I think it isn't true as it is too early for it to be real.
.....
Then your feeling how 3070Ti will trade blows with Big Navi is most likely wrong.
80CU Big Navi can hardly perform only on the level of 3070Ti which is just a bit faster than 2080Ti according to that graph. What's interesting in that graph is that Ampere has only a bit better perf/W than Turing without RT. Most likely this graph is not correct, but at least we can discuss something.
 
  • Like
Reactions: Saylick and Konan

Glo.

Diamond Member
Apr 25, 2015
5,803
4,777
136
I don't think people who are making that graph even grasp how big IPC increase would Nvidia need to get for RTX 3070 with 40 SMs(2560 ALUs) to be faster than RTX 2080 Super with 48 SM's(3072 ALUs). Not to mention 48 SM GPU, to be faster than 72 SM GPU from Turing architecture...? KEK.

Remember. RTX 3080 is based on 102 die, and has 68 SM's. Next is 104 die, with 48 SM. And that is 3070 Ti.

Also the gap in performance between 48 SM and 68 SM GPU is way too small, which makes this graph even worse.
 

jpiniero

Lifer
Oct 1, 2010
15,223
5,768
136
Also the gap in performance between 48 SM and 68 SM GPU is way too small, which makes this graph even worse.

Remember frequency. The mentioned frequency might be technically the set boost but it can go much higher than that in games. Chart at first glance doesn't seem that out of place but would need to look at it more detailed.
 

Konan

Senior member
Jul 28, 2017
360
291
106
I don't think people who are making that graph even grasp how big IPC increase would Nvidia need to get for RTX 3070 with 40 SMs(2560 ALUs) to be faster than RTX 2080 Super with 48 SM's(3072 ALUs). Not to mention 48 SM GPU, to be faster than 72 SM GPU from Turing architecture...? KEK.

Remember. RTX 3080 is based on 102 die, and has 68 SM's. Next is 104 die, with 48 SM. And that is 3070 Ti.

Also the gap in performance between 48 SM and 68 SM GPU is way too small, which makes this graph even worse.
GA104 is rumored to have 48SMs by KatKorgi and 3072 Cuda cores. Add in IPC, frequency and FP32 performance changes.
Or like they said you take TSE scores and work back from that using the differences they say.... the chart looks plausible in places

Edit:

GA104 as per KatCorgi within 5% of 2080Ti it is right there what was said back from March in the thumbnail. And that could be just the plain old 3070 and not the Ti version that is also now rumored in a few places.
What has changed since then? Rumor apparently says with 16GB now to help with greater FP32 throughput and now we have GDDR6X confirmation.

1597597557991.png
 
Last edited:

Konan

Senior member
Jul 28, 2017
360
291
106
Then your feeling how 3070Ti will trade blows with Big Navi is most likely wrong.
80CU Big Navi can hardly perform only on the level of 3070Ti which is just a bit faster than 2080Ti according to that graph. What's interesting in that graph is that Ampere has only a bit better perf/W than Turing without RT. Most likely this graph is not correct, but at least we can discuss something.

Well CoreTeks was pitching the BN performance at a little over the 2080Ti.

Nice point on the perf/w observation. Agree it is impossible for someone to know all the performance numbers at this point 100%.
 

Glo.

Diamond Member
Apr 25, 2015
5,803
4,777
136
Remember frequency. The mentioned frequency might be technically the set boost but it can go much higher than that in games. Chart at first glance doesn't seem that out of place but would need to look at it more detailed.
2100 MHz is the core clock limit. It won't go past that.
 

jpiniero

Lifer
Oct 1, 2010
15,223
5,768
136
2100 MHz is the core clock limit. It won't go past that.

That might be enough. Say the 3070 Ti FE actually hits that... same shader count as 2080 Super, and it's 20% faster. 9% frequency boost plus 10% IPC gain gets you pretty close.

Note I do not expect the 3070 Ti to be released any time soon. The 3070 was earlier rumored to be 2944 (same as 2080) but that could change.
 

TESKATLIPOKA

Platinum Member
May 1, 2020
2,523
3,037
136
GA104 is rumored to have 48SMs by KatKorgi and 3072 Cuda cores. Add in IPC, frequency and FP32 performance changes.
FP32 performance changes? What's that supposed to mean? Isn't that actually better IPC?
GA104 has supposedly 3072 FP32 Cuda cores in 48SMs(64 Cuda per SM as with Turing). 2080Ti has 68SMs active so the difference is 42%. RTX 2080Ti FE has on average 1824Mhz so If I set the clockspeed for GA104 10-15% higher that would mean 2006-2098Mhz. It leaves 23-29% for better IPC to be on par with 2080Ti. A bit high in my opinion.
Personally I think RDNA2 will have higher clocks than Ampere. Why do I think so? Because Renoir IGP is capable of 2.4-2.5Ghz OC and that's on old Vega architecture and no Vega GPU was capable of that before. Let's not forget about PS5 GPU turbo clockspeed of 2.23Ghz.
 
Last edited:

jpiniero

Lifer
Oct 1, 2010
15,223
5,768
136
FP32 performance changes? What's that supposed to mean? Isn't that actually better IPC?
GA104 has supposedly 3072 FP32 Cuda cores in 48SMs(64 Cuda per SM as with Turing). 2080Ti has 68SMs active so the difference is 42%. RTX 2080Ti FE has on average 1824Mhz so If I set the clockspeed for GA104 10-15% higher that would mean 2006-2098Mhz. It leaves 23-29% for better IPC to be on par with 2080Ti. A bit high in my opinion.
Personally I think RDNA2 will have higher clocks than Ampere. Why do I think so? Because Renoir Vega IGP is capable of 2.4-2.5Ghz OC.

2080 Ti might be memory bottlenecked.