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 know we've had a lot of speculation and even a few leaks of varying degrees of believability, but we don't have final prices yet.

They could certainly change it, but the rumors are saying $799 for the 3080 and $1399 for the 3090, with the FE being $100 more. No reason to doubt that is accurate.

 

[raghu78]

Yeah, I'm not surprised that Nvidia even made that video because they are the king of marketing after all. They are selling the concept as "Our goals was to push as much power, and thus performance, out of our cards as possible, and so we had to jump through extra hoops to make it work". Average consumer probably sees that as, "Oh cool, Nvidia want to give us more performance no matter the cost and had to design a new cooler to enable that" while others like myself see it as "Oh, we couldn't improve the perf/W high enough so that our new generation of cards have a big enough performance delta over the last generation, so we had to crank up the clocks and power to get there." I have no doubt in my mind that if Nvidia sourced TSMC N7P for Ampere consumer GPUs, they wouldn't have been in this dilemma. In my mind, there is easily a 20% perf/W improvement between SS 8nm and TSMC N7P, or enough where the RTX 3090 could have been a 300W card, not 350W.

TSMC N7P/N7+ is 15-18% higher transistor perf than Samsung 8LPP. At the same transistor performance the power is going to be lower by 30-35%. If Nvidia had made GA102 and the rest of the stack on TSMC N7+ they would not be struggling with power. In fact Nvidia would likely have pushed all the way to 96 SM and got excellent perf, perf/sq mm and perf/watt. I would not have been surprised to see another Maxwell to Pascal jump if Nvidia had fabbed Ampere Geforce on TSMC N7+.

 

[n0x1ous]

is the move to Samsung about supply, price, or both? JHH did say that vast majority of Ampere was going to TSMC. Maybe with Datacenter overtaking GeForce in revenue now he means in dollars sense and HPC/datacenter is all TSMC and consumer/pro is Samsung. With AMD now using so many TSMC wafers after GF gave up on shrinks after 14nm maybe Nvidia really felt they couldn't get enough supply. next week will be interesting.

 

[Glo.]

is the move to Samsung about supply, price, or both? JHH did say that vast majority of Ampere was going to TSMC. Maybe with Datacenter overtaking GeForce in revenue now he means in dollars sense and HPC/datacenter is all TSMC and consumer/pro is Samsung. With AMD now using so many TSMC wafers after GF gave up on shrinks after 14nm maybe Nvidia really felt they couldn't get enough supply. next week will be interesting.

JHH said vast majority of 7 nm products is going to TSMC .

 

[n0x1ous]

JHH said vast majority of 7 nm products is going to TSMC .

Since Samsung is 8nm he gets away with this statement I suppose then.

 

[Glo.]

So what then? No way he straight up lied to investors about that. Is all the samsung talk a farce?

Do you understand the difference between 8 nm process from Samsung, and 7 nm process from TSMC?

He said that majority of 7 nm products are made on TSMC.

Gaming next gen cards are 8 nm Samsung, so why would he lie about anything?

He stated the true fact, actually. If the only 7 nm product is A100 chip, then it is made on TSMC's process.

Why complicate simple things?

 

[jpiniero]

In fact Nvidia would likely have pushed all the way to 96 SM and got excellent perf, perf/sq mm and perf/watt. I would not have been surprised to see another Maxwell to Pascal jump if Nvidia had fabbed Ampere Geforce on TSMC N7+.

I believe that GA102 was always intended to be 84 SM, even on SS7, and they seem to have hit the targeted frequency on SS8. Of course it's drawing decently more power.

 

[Stuka87]

is the move to Samsung about supply, price, or both? JHH did say that vast majority of Ampere was going to TSMC. Maybe with Datacenter overtaking GeForce in revenue now he means in dollars sense and HPC/datacenter is all TSMC and consumer/pro is Samsung. With AMD now using so many TSMC wafers after GF gave up on shrinks after 14nm maybe Nvidia really felt they couldn't get enough supply. next week will be interesting.

nVidia tried playing hardball with TSMC to get them to drop their prices by announcing they were going with Samsung. TSMC said "later, we aren't changing our prices." So then nVidia was stuck with Samsung, but knew they couldn't use Samsung for A100. So they managed to get some wafers from TSMC for the A100 chips. But since TSMC had already sold the vast majority of their 7nm capacity to tier 1 and 2 customers (AMD, Qualcomm, Broadcom, Apple) nVidia was stuck with the left overs.

Which leads us to why all their consumer GPU's are based on the Samsung process.

 

[uzzi38]

Since Samsung is 8nm he gets away with this statement I suppose then.

He still doesn't get let off his statement that a 2080 Max-Q laptop will outperform the next gen consoles if you ask me.

 

[Saylick]

nVidia tried playing hardball with TSMC to get them to drop their prices by announcing they were going with Samsung. TSMC said "later, we aren't changing our prices." So then nVidia was stuck with Samsung, but knew they couldn't use Samsung for A100. So they managed to get some wafers from TSMC for the A100 chips. But since TSMC had already sold the vast majority of their 7nm capacity to tier 1 and 2 customers (AMD, Qualcomm, Broadcom, Apple) nVidia was stuck with the left overs.

Which leads us to why all their consumer GPU's are based on the Samsung process.

This, and also that Samsung over-hyped and under-delivered on their process. If Nvidia had known that they'd be fabbing on such an inferior process from the get-go, I'm not so sure they would have went down the route of using Samsung, even if Samsung had pitched to Nvidia that their wafers were 2/3rds of the cost of TSMC's.

 

[Ajay]

This, and also that Samsung over-hyped and under-delivered on their process. If Nvidia had known that they'd be fabbing on such an inferior process from the get-go, I'm not so sure they would have went down the route of using Samsung, even if Samsung had pitched to Nvidia that their wafers were 2/3rds of the cost of TSMC's.

True, but Nvidia *really* likes cheap wafers. They've been a node behind AMD for a number of GPUs.

 

[Tup3x]

He still doesn't get let off his statement that a 2080 Max-Q laptop will outperform the next gen consoles if you ask me.

Well, there's Xbox Series S (Lockhart) coming so... At least that one.

 

[lobz]

He still doesn't get let off his statement that a 2080 Max-Q laptop will outperform the next gen consoles if you ask me.

Well that is most certainly not true.

 

[Stuka87]

Well, there's Xbox Series S (Lockhart) coming so... At least that one.

Even then its arguable. As laptop solutions tend to suffer from cooling issues. Fast benchmarks look fine. But play for 30 mins, and things tend to throttle. But comparing console performance and PC performance tends to be tough. It has to be generalized to "yes the PC can have similar quality settings and sustain 60fps at a given resolution".

 

[sontin]

TSMC N7P/N7+ is 15-18% higher transistor perf than Samsung 8LPP. At the same transistor performance the power is going to be lower by 30-35%. If Nvidia had made GA102 and the rest of the stack on TSMC N7+ they would not be struggling with power. In fact Nvidia would likely have pushed all the way to 96 SM and got excellent perf, perf/sq mm and perf/watt. I would not have been surprised to see another Maxwell to Pascal jump if Nvidia had fabbed Ampere Geforce on TSMC N7+.

Navi is on 7nm and failed against Pascal. Maybe we should think about that. And yet we have people here who are claiming that AMD will make Navi 2 more than twice as efficient as Navi 1. How should this work?

 

[n0x1ous]

navi 2 has to be almost all architectural gains since they were already on 7nm for navi 1. Nvidia has any architectural improvements plus the full node jump. No way navi 2 wins against ampere unless Nvidia royally screwed up. Which judging by the last 3 or 4 rounds is highly unlikely

 

[jpiniero]

Navi is on 7nm and failed against Pascal. Maybe we should think about that. And yet we have people here who are claiming that AMD will make Navi 2 more than twice as efficient as Navi 1. How should this work?

RDNA1 was badly broken and they fixed it? Plus N7P instead of N7.

 

[Glo.]

navi 2 has to be almost all architectural gains since they were already on 7nm for navi 1. Nvidia has any architectural improvements plus the full node jump. No way navi 2 wins against ampere unless Nvidia royally screwed up. Which judging by the last 3 or 4 rounds is highly unlikely

If you would pay attention, you would find that AMD achieved all: higher clock speeds, higher IPC, and better efficiency, all at the same time.

Look at Xbox Series X GPU. 52 CU's clocked at 1.8 GHz for 130-140W of power. PS5 clocks up to 2.23 GHz.

And yes, Nvidia screwed up, by completely underestimating AMD's capability of delivering good products.

Why being affraid of this reality?

Thirdly. Its Nvidia thread...

 

[Glo.]

Navi is on 7nm and failed against Pascal. Maybe we should think about that. And yet we have people here who are claiming that AMD will make Navi 2 more than twice as efficient as Navi 1. How should this work?

You don't have to fight with reality.

Especially since we have Xbox Series X, and PS5 data to do the math.

 

[Tup3x]

Even then its arguable. As laptop solutions tend to suffer from cooling issues. Fast benchmarks look fine. But play for 30 mins, and things tend to throttle. But comparing console performance and PC performance tends to be tough. It has to be generalized to "yes the PC can have similar quality settings and sustain 60fps at a given resolution".

If it's 4 teraflops like in rumours then it doesn't take much to beat it.

 

[DDH]

Navi is on 7nm and failed against Pascal. Maybe we should think about that. And yet we have people here who are claiming that AMD will make Navi 2 more than twice as efficient as Navi 1. How should this work?

How did Navi fail against Pascal? Spec wise 5700xt and 2070 super are almost identical, so offer the best case for a comparison. Their performance are within a few % of each other most of the time, and clock for clock 5700xt uses less power

In a single generation AMD has matched the IPC of Turing and beaten it in efficiency, albeit with a node advantage.

Nevertheless, if AMD is correct in an improvement of 50% performance per watt, and the power envelope of the Xbox series X seems to indicate that they are, then imagine the above video with the 5700xt pulling 120 - 130w.

Sent from my SM-N975F using Tapatalk

 

[n0x1ous]

I'd love for Navi 2 to be competitive or even better, but I'm used to being disappointed by AMD gpu's after owning Fury X, Vega 64, and Radeon VII

 

[sontin]

RDNA1 was badly broken and they fixed it? Plus N7P instead of N7.

AMD already uses N7P for Navi.

 

[DDH]

AMD already uses N7P for Navi.

No, Navi 10 uses N7FF

Sent from my SM-N975F using Tapatalk

 

[Krteq]

No, Navi 10 uses N7FF

Source?