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.

 

[kurosaki]

From the latest Techpowerup review, the RTX2080ti is 40%-50% faster vs the RX5700XT at 1440p and 4K.
If RTX3090 is 50% faster vs RTX2080Ti then AMD needs to have a 100% faster card vs the RX5700XT to get even, not 225%

This. The AMD 6xxx-series is going to be good enough for most people if priced right. Just look at Valves yearly Steam-graphs. There are very few high end cards on those lists and are completely dominated by the best bang for the buck mid tiers. If AMD isn't going all Nvidia on the pricing this year, we have exciting times ahead. Nvidia is bound to try to compete in pricing and hopefully we will finally get cards that can run 1440p with RT and with high enough fps at the same time. For less than the cost of a used car, or a small internal organ.

 

[Glo.]

This. The AMD 6xxx-series is going to be good enough for most people if priced right. Just look at Valves yearly Steam-graphs. There are vere few high end cards on those lists and are completely dominated by the best bang for the buck mid tiers. If AMD isn't going all Nvidia on the pricing this year, we have exciting times ahead. Nvidia is bound to try to compete in pricing and hopefully we will finally get cards that can run 1440p with RT and with high enough fps at the same time. For less than the cost of a used car, or a small internal organ.

It depends on Nvidia pricing. Don't expect that big Navi if competitive to be priced Below 1000$. Don't expect that Small Navi, Navi 23 to be priced below RTX 3070, if it is competitive with it.

Hasn't this generation taught us nothing? AMD won't play the value brand anymore if they have premium products. So our only hope for lower prices is actually - Nvidia.

 

[kurosaki]

It depends on Nvidia pricing. Don't expect that big Navi if competitive to be priced Below 1000$. Don't expect that Small Navi, Navi 23 to be priced below RTX 3070, if it is competitive with it.

Hasn't this generation taught us nothing? AMD won't play the value brand anymore if they have premium products. So our only hope for lower prices is actually - Nvidia.

Well, that's how you win over customers, if you have card to sell that is. I'm afraid that TW and Samsung will have problems keeping volumes up.

 

[Glo.]

Well, that's how you win over customers, if you have card to sell that is. I'm afraid that TW and Samsung will have problems keeping volumes up.

Well, I wouldn't mind RTX 3060 for 299$ if it had RTX 2070 Super-RTX 2080 performance .

 

[kurosaki]

Well, I wouldn't mind RTX 3060 for 299$ if it had RTX 2070 Super-RTX 2080 performance .

3060? Your aim is to low! 3070s for 300usd, 3080(+) can go for another 150. Hold up the buying until the sales are raining in!

 

[Glo.]

Guys, seriously, the more I think about it, the more I am baffled, how come people have expectations about Ampere performing vastly more powerful than Turing in games?

107 die - 24 SM's, 106 die, 36 SM's, 104 die - 48 SM's.

107 die - 128 bit bus, 106 die 192 bit bus, 104 die - 256 bit bus.

Clock targets - up to, what, 2.1-2.15 Ghz?

Where do people get the idea that 48 SM/3072 ALU GPU, with only 10% higher clock speed than RTX 2080 Super, with at best 20% more memory bandwdith, will be faster(!) than RTX 2080 Ti?

3060? Your aim is to low! 3070s for 300usd, 3080(+) can go for another 150. Hold up the buying until the sales are raining in!

Nah, man. It won't happen.

 

[kurosaki]

Guys, seriously, the more I think about it, the more I am baffled, how come people have expectations about Ampere performing vastly more powerful than Turing in games?

107 die - 24 SM's, 106 die, 36 SM's, 104 die - 48 SM's.

107 die - 128 bit bus, 106 die 192 bit bus, 104 die - 256 bit bus.

Clock targets - up to, what, 2.1-2.15 Ghz?

Where do people get the idea that 48 SM/3072 ALU GPU, with only 10% higher clock speed than RTX 2080 Super, with at best 20% more memory bandwdith, will be faster(!) than RTX 2080 Ti?

Nah, man. It won't happen.

Just like the good old days, performance per dollar doubling every year...

 

[jpiniero]

Guys, seriously, the more I think about it, the more I am baffled, how come people have expectations about Ampere performing vastly more powerful than Turing in games?

107 die - 24 SM's, 106 die, 36 SM's, 104 die - 48 SM's.

107 die - 128 bit bus, 106 die 192 bit bus, 104 die - 256 bit bus.

Clock targets - up to, what, 2.1-2.15 Ghz?

Where do people get the idea that 48 SM/3072 ALU GPU, with only 10% higher clock speed than RTX 2080 Super, with at best 20% more memory bandwdith, will be faster(!) than RTX 2080 Ti?

IPC gains of course. Difference between 2080S and 2080 Ti FE is about 20%.

 

[Qwertilot]

Guys, seriously, the more I think about it, the more I am baffled, how come people have expectations about Ampere performing vastly more powerful than Turing in games?

Because they've, one way or another, basically delivered the same levels of annual performance increases for a decade or so? Its like clockwork.

No real reason to think its suddenly entirely snarled up with a decent die shrink to work with, and a good while to do the new architecture.

Honestly, it makes all this chasing after rumoured specifications a bit futile. We more or less know what we're getting in advance
(How they manage it each time is a bit different.).

 

[Glo.]

IPC gains of course. Difference between 2080S and 2080 Ti FE is about 20%.

Because they've, one way or another, basically delivered the same levels of annual performance increases for a decade or so? Its like clockwork.

No real reason to think its suddenly entirely snarled up with a decent die shrink to work with, and a good while to do the new architecture.

Honestly, it makes all this chasing after rumoured specifications a bit futile. We more or less know what we're getting in advance
(How they manage it each time is a bit different.).

You do realize that with Turing ALL of IPC gains came from... Increased memory bandwdith?

Compare GTX 1650 with GTX 1050 Ti. All of performance in GTX 1650 came from increased ALU count, and higher clock speeds, sustained.

So let me ask you this. Out of what Nvidia is able to squeeze more IPC gains, without vast redesign of their architecture?

Magic? Kryptonite? Adamantium?

Secondly. What has past to do with the future. Intel has been on a good proces execution for past 20 years so its impossible for them to NOT deliver next node, eh? AMD failed to execute on CPU side for past 10 years so they will never come back, eh?

"Look where the puck is going, not where it has been."

 

[tviceman]

From the latest Techpowerup review, the RTX2080ti is 40%-50% faster vs the RX5700XT at 1440p and 4K.
If RTX3090 is 50% faster vs RTX2080Ti then AMD needs to have a 100% faster card vs the RX5700XT to get even, not 225%

If the 3090 is "50%" faster than the 2080 TI, then it will be 125% faster than the 5700 XT.

1.5 x 1.5 = 2.25

Conversely, if Big Navi is 2x faster than 5700 XT, then it will be 33% faster than the 2080 TI.

0.667 x 2.0 = 1.334

But come on, a new node, a new architecture, and a new crazy high TDP - only for 3090 to be 50% faster than the 2080 TI? I don't buy that.

If the RTX 3090 Ampere is a conservative 30% more efficient than RTX 2080 TI Turing, that alone will translates into a 65% performance uplift over Turing based on the rumored TDP. So then 1.5 x 1.65 = 2.475, which is nearly 150% faster than the 5700 XT. But this is just at rasterization. Nvidia may (or may not) also hold a very commanding lead in RT capabilities.

 

[maddie]

This one is huuuge

WCCFTech - NVIDIA GeForce RTX 3090 & RTX 3080 Ampere GA102 GPU Allegedly Pictured – Massive Die For Enthusiast Gaming Graphics Cards

That link allowed me to see the admittedly leaked, 3080 & 3090 card sizes. IF the 3080 is only a cut down 102 die (68 SM) and the 3090 (82 SM) is the full one, then why in the world would you need around a 100% increase in cooler volume?

3090 cooler = 125-140% greater area and 3 slots versus 2 slots. This further shows me, though I realize I'm swimming against the tide here, that I really see a dual GPU card here.

 

[Gideon]

From the latest Techpowerup review, the RTX2080ti is 40%-50% faster vs the RX5700XT at 1440p and 4K.
If RTX3090 is 50% faster vs RTX2080Ti then AMD needs to have a 100% faster card vs the RX5700XT to get even, not 225%

Sorry for being unclear. I meant 225% of the absolute perfomance of RX 5700XT not 225% faster. That would be 125% faster.

But Basic 6th-grade math?

If RX 5700XT gets 1000 points in an imaginary benchmark. RTX 2080 Ti gets 1500 points. Now if RTX 3090 is 50% faster than RTX 2080 Ti what will it's score be? (hint: it's not 2000 )

 

[tviceman]

That link allowed me to see the admittedly leaked, 3080 & 3090 card sizes. IF the 3080 is only a cut down 102 die (68 SM) and the 3090 (82 SM) is the full one, then why in the world would you need around a 100% increase in cooler volume?

3090 cooler = 125-140% greater area and 3 slots versus 2 slots. This further shows me, though I realize I'm swimming against the tide here, that I really see a dual GPU card here.

I imagine it's because the memory voltage @ 2.1 ghz and the clock speeds being pushed hard on the fully unlocked chip. Look at the notebook versions of Turing - they are able to get 90% of the performance of their desktop equivalents with a 33% drop in TDP. That is what the 3080 will do to get it's TDP to sane (yet still high) levels.

 

[tviceman]

Sorry for being unclear. I meant 225% of the absolute perfomance of RX 5700XT not 225% faster. That would be 125% faster.

But Basic 6th-grade math?

If RX 5700XT gets 1000 points in an imaginary benchmark. RTX 2080 Ti gets 1500 points. Now if RTX 3090 is 50% faster than RTX 2080 Ti what will it's score be? (hint: it's not 2000 )

Haha I beat you to it a few posts above.

 

[AtenRa]

If the 3090 is "50%" faster than the 2080 TI, then it will be 125% faster than the 5700 XT.

yea my bad, its 125% not 100%

 

[maddie]

I imagine it's because the memory voltage @ 2.1 ghz and the clock speeds being pushed hard on the fully unlocked chip. Look at the notebook versions of Turing - they are able to get 90% of the performance with a 33% drop in TDP. That is what the 3080 will do to get it's TDP to sane (yet still high) levels.

Twice the power dissipation capacity available? Don't see it as due to memory and OC.

 

[AtenRa]

Sorry for being unclear. I meant 225% of the absolute perfomance of RX 5700XT not 225% faster. That would be 125% faster.

But Basic 6th-grade math?

If RX 5700XT gets 1000 points in an imaginary benchmark. RTX 2080 Ti gets 1500 points. Now if RTX 3090 is 50% faster than RTX 2080 Ti what will it's score be? (hint: it's not 2000 )

yes my bad, i fixed to 125%

 

[xpea]

You do realize that with Turing ALL of IPC gains came from... Increased memory bandwdith?

So let me ask you this. Out of what Nvidia is able to squeeze more IPC gains, without vast redesign of their architecture?

Magic? Kryptonite? Adamantium?

Secondly. What has past to do with the future. Intel has been on a good proces execution for past 20 years so its impossible for them to NOT deliver next node, eh? AMD failed to execute on CPU side for past 10 years so they will never come back, eh?

"Look where the puck is going, not where it has been."

You do realize that with RDNA2 ALL of IPC gains came from... Increased memory bandwdith?

So let me ask you this. Out of what RDNA2 is able to squeeze more IPC gains, without vast redesign of their architecture?

Magic? Kryptonite? Adamantium?

 

[trexfromouterspace]

XFX had better get back to manufacturing Nvidia cards so they can gift us the RTX 3090 BIG CHUNGUS XTREME

 

[Glo.]

You do realize that with RDNA2 ALL of IPC gains came from... Increased memory bandwdith?

So let me ask you this. Out of what RDNA2 is able to squeeze more IPC gains, without vast redesign of their architecture?

Magic? Kryptonite? Adamantium?

Caches redesigned, redesigned scheduling, More instructions executed, per clock, higher memory bandwidth, higher cache bandwidth. Is it enough?

We know what Ampere is and apart from GEMM stuff, its the same Turing architecture, on this very front.

So let me ask again. Where do all of those IPC improvements are going to come from?

 

[Gideon]

So let me ask you this. Out of what Nvidia is able to squeeze more IPC gains, without vast redesign of their architecture?

Magic? Kryptonite? Adamantium?

What makes you think the architecture is unchanged? There are rumors of Nvidia doubling FP32 perfomance and they've managed to improve their perf/watt before on the same node (take Maxwell for instance).

Let's take a trip down memory lane to the first page of this thread:

An NVIDIA employee told me I'll be impressed...that's all I got.

They should've known fairly well what the performance targets will be in November last year (they most certainly knew the node they were targeting). I doubt extracting no extra IPC gain with the exact same architecture would count as "impressive".

We will find out soon-enough. I wouldn't mind at all if RDNA2 would be half as good in rasterizing perfomance as you claim.

 

[tviceman]

Twice the power dissipation capacity available? Don't see it as due to memory and OC.

It's a $1,300+ product. Nvidia has a penchant for using premium components (and then charging for it). It won't be a dual-GPU card. Those days are gone until MCM makes that solution seemless.

 

[Glo.]

What makes you think the architecture is unchanged? There are rumors of Nvidia doubling FP32 perfomance and they've managed to improve their perf/watt before on the same node (take Maxwell for instance).

Erm, we know what Ampere is, and apart from GEMM stuff, its the same Turing architecture on the front of caches, scheduling throughput, etc.

If gaming cards are different on this front, why Nvidia is calling it Ampere?

P.S. Don't get me wrong. I won't exclude that Nvidia somehow increased IPC on next gen gaming cards compared to Turing. But it won't be as huge as everybody hopes for, especially in order for it to achieve 60% higher rasterization performance than Turing.

And based on what we know about Ampere architecture already - there might not be any IPC increase on the ALU level.

 

[Gideon]

Where do people get the idea that 48 SM/3072 ALU GPU, with only 10% higher clock speed than RTX 2080 Super, with at best 20% more memory bandwdith, will be faster(!) than RTX 2080 Ti?

You do realize that with Turing ALL of IPC gains came from... Increased memory bandwdith?

This statement is incorrect (at least when made universally).
Looking at techpowerup RTX 2080 review (unfortunately Anandtech bench doesn't have 1080 ti):

• RTX 2080 is 8% faster (on average) than GTX 1080 Ti in every resolution
• That is despite having 8% less memory bandwidth, 18% less shader-cores and only a 8% faster boost clock to compensate (shaders more than offset the difference).

On equal terms it would be over 15% faster.