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.

 

[sontin]

Its not a matter of making Nvidia bleed on high end. Current picture of the information we got on hand says that AMD will have better chips, from top, to bottom, with new generation. The only place where they might not have better products is... low-end. 107 and 106 might be better than... Navi 10 based GPUs.

How exactly can AMD have "better chips" when nVidia has the advantage of a one or two full node shrinks? Navi failed against Pascal. Now think about Navi 2 against Ampere with twice the transistors.

 

[AtenRa]

RDNA 2 has RTRT, but AMD went about it in a very different way. Each shader had the functionality built into it. Rather than having a big chunk of the GPU dedicated to hardware that can only do RTRT. What this means for performance, we don't know. But in theory, it means the AMD GPU's could be more balanced. Which was one of Turing's failings. Not nearly enough RT hardware to let the card run at regular rasterized speeds.

I know RDNA2 has hardware RT support, I was only pointing out that i believe they didnt invested as much silicon to RT and especially tensors like NV did in Ambere.

 

[uzzi38]

Let's see them do it once without a full node drop.

"Poor Volta"

Three years waiting on that claim, but now they are trustworthy?

Like I said in that post, RDNA2 is the proving ground. You haven't got long to wait now, hell we may even get a tease next week.

There's one more announcement from AMD on the 28th after all.

 

[jpiniero]

re: 300 W, the official spec for the 3080 Ti might be 300 W but the power draw being talked about is what the FE is rated for, which is probally closer to 350.

Now think about Navi 2 against Ampere with twice the transistors.

Twice the transistors compared to Pascal maybe, not Turing.

 

[Glo.]

Now think about Navi 2 against Ampere with twice the transistors.

This sentence is not going to age well.

For example. TU102 has 18 bln transistors. Rumored die size for 102 die from next gen is 627 mm2.

In order to achieve 2x transistor count, you would need 55-60 mln xTors/mm2.

Which is absolutely impossible for this node. We are talking about 40 mln xTors/mm2 for large chips, with high clock speeds, in absolutely best case scenario. 35 mln xTors/mm2 is more realistic, for 8 nm SS. Remember, this is slightly tweaked 10 nm process, in similar way to 12 FFN was slightly tweaked 16 nm TSMC process. Its more similar to TSMC's 10 nm process, which is one process node ago.

 

[traderjay]

Maybe its for the 3090 ti a dual gpu monster

 

[Veradun]

Let's see them do it once without a full node drop.

"Poor Volta"

Three years waiting on that claim, but now they are trustworthy?

I'm pretty sure they sold a lot more consumer Vegas than nvidia did with consumer Voltas

 

[Veradun]

RDNA 2 has RTRT, but AMD went about it in a very different way. Each shader had the functionality built into it.

Isn't it in the TMUs?

 

[itsmydamnation]

Isn't it in the TMUs?

no, it just uses the TMU datpaths.

 

[Krteq]

Isn't it in the TMUs?

Yes, "Intersection Engine" is a part of TMU according to patent

Spoiler: Patent schemas

no, it just uses the TMU datpaths.

Nope, "Intersection Engine" has the same basic functionality as RT core, it accelerates ray intersection testing in BVH structure (see FIG. 18 in spoiler above)

 

[TESKATLIPOKA]

How exactly can AMD have "better chips" when nVidia has the advantage of a one or two full node shrinks? Navi failed against Pascal. Now think about Navi 2 against Ampere with twice the transistors.

Navi failed against Pascal? Then I don't understand why Nvidia had to release Super cards based on Turing If Navi according to you failed even against Pascal.

Back to the topic.
I would rather wait to see what will come out of Ampere, but I don't expect a bad chip from Nvidia, they delivered good products after the Fermi architecture.
It looks like AMD will finally release a much bigger chip compared to the previous generation and I consider this the biggest problem of RDNA1 -> relatively small chips. This time I expect a closer fight between AMD and Nvidia, but I don't dare claim which architecture will be better. I would be more than happy If the price/performance improves at least by 25%.

 

[AtenRa]

Navi failed against Pascal? In what did It fail? Please enlighten me.

Yea im also trying hard to find out where exactly Navi failed against Pascal.

 

[sontin]

Navi failed against Pascal? In what did It fail? Please enlighten me.

10,2 billions xtors (32% more) and only ~20% faster with ~25% higher power consumption than GTX1080. Navi is already on 7nm. And now people claiming here that nVidia would have no chance with Ampere.

 

[AtenRa]

I can also see NAVI 10 (RX5700XT) with 1.5Billion fewer transistors being as fast as GP102 GTX1080Ti

 

[sontin]

I can also see NAVI 10 (RX5700XT) with 1.5Billion fewer transistors being as fast as GP102 GTX1080Ti

Sure, if you ignore more than 90% of the games. Full GP102 is 20%+ faster.

 

[amenx]

I can also see NAVI 10 (RX5700XT) with 1.5Billion fewer transistors being as fast as GP102 GTX1080Ti

So took them 2.5 years to achieve parity with Pascal performance.

 

[AtenRa]

So took them 2.5 years to achieve parity with Pascal performance.

If by your own math 1.5B fewer transistors is achieving parity, then yes.

 

[AtenRa]

Sure, if you ignore more than 90% of the games. Full GP102 is 20%+ faster.

I dont even have to comment on that

 

[TESKATLIPOKA]

Double post.

 

[TESKATLIPOKA]

10,2 billions xtors (32% more) and only ~20% faster with ~25% higher power consumption than GTX1080. Navi is already on 7nm. And now people claiming here that nVidia would have no chance with Ampere.

Based on this Link and this Link
5700XT vs 1080 is ~27% faster and power consumption is 32% higher, transistor count is 43% higher(10.3 vs 7.2 billion transistors).
Let's compare 5700 vs 1080
5700XT vs 1080 is ~13% faster and power consumption is the same, transistor count is 43% higher(10.3 vs 7.2 billion transistors).
It's true they used too much transistors for not so high performance gain so from this point It could be considered as "fail" and power consumption is also not so good considering the advantage of 7nm.
People claiming how Ampere will stand no chance against Navi2 is pure hype based on too little known info. Nvidia would need to f**k up big time for this to happen.

 

[TESKATLIPOKA]

I dont even have to comment on that

He was talking about Nvidia TITAN Xp and not the cut down 1080Ti, but even so the difference in performance wouldn't be 20% but more likely 10-15% and considering It has only 16.5% more transistors then yes they achieved parity after 2.5 years. Let's be honest, Pascal was and still is a great architecture.

 

[uzzi38]

Btw not sure if it's been posted yet but the 12 pin is indeed real and it's designed by Nvidia:

 

[Stuka87]

10,2 billions xtors (32% more) and only ~20% faster with ~25% higher power consumption than GTX1080. Navi is already on 7nm. And now people claiming here that nVidia would have no chance with Ampere.

Sure, if you ignore more than 90% of the games. Full GP102 is 20%+ faster.

Dude, on what planet? First, Titan XP is single digits faster than a 1080Ti. In many games, the AIB 1080Ti's are FASTER than Titan XP. The few other cuda cores that Titan XP has don't help it in games at all.

The "90% of games" remark also carries no weight. Just look at TPU's results, which are always skewed in nVidia's favor regarding game selection. The 1080Ti is a wopping 4% faster than a 5700XT. And thats with the 5700XT being significantly smaller and cheaper. ANNNND... The 5700 XT (per TPU) uses LESS POWER than a 1080Ti. With the 5700XT using 219W and the 1080Ti using 231W. If we get into AIB cards power usages and speeds jump around for both. This is reference vs reference.

 

[Konan]

What people here don't seem to realise is that it's not RDNA2 that's the real danger to Nvidia. The next generation will be close but that's all. It's AMD's swift cadence and +50% perf/W each generation promise that is a real danger.

A cadence shift like this is not easy at all.

Been there. Heard it before.

Retrospective quote from PC Gamer:

There are many changes with Navi, the first new GPU architecture for AMD graphics cards since its RX Vega line in 2017, but after the official reveal of specs and pricing at E3 2019, I have to wonder if this is truly new or simply the next iteration on the existing product line. After many years of playing second string to Nvidia's leading parts, I wanted a clear win from AMD. Navi, sadly, isn't it. AMD claims IPC performance improvements of 25 percent per CU, and says that overall performance per watt will be 50 percent better than its previous generation Vega and Polaris architectures.

That sounds great, but when you dig into the details, it looks like at best AMD might match the performance of Nvidia's RTX 2060 and 2070 GPUs, except AMD will use more power, potentially cost a bit more, and Navi doesn't include any ray tracing or deep learning features. Considering AMD is using TSMC's latest 7nm manufacturing process, it's a bit of a letdown—a lot like the Fiji, Polaris, and Vega GPUs of the past several years.

I wonder how much is the same for this forthcoming time around.
This “+50% perf/W each generation promise“ Sounds great and is definitely a decent KPI to aim for. But the reality has been hasn’t really gotten them anywhere at the medium/high to high end. No Ray tracing, no deep learning features. They are way behind on software + SDK. I have some serious confidence issues but I hope to be proven wrong on the software side with.
Navi 1x was supposed to launch early 2018 it was late, way late. I don’t believe any capacity or capability to move any quicker from them based on demonstrated lateness at every new node or uArch change. Also Navi 2X does not mean 2*5700XT.

 

[Konan]

Apparently at this point in time... the Engineering Sample RTX 3080 these guys see show 20% + performance over a RTX 2080Ti and the final version a bit more, say 22-25% I think if that is what they mean.
So rumoured TGP of what, 300-320W for up to 25% more (3080>2080Ti) that may put a potential RTX 3070 between a 2080 Super and 2080Ti then?

https://twitter.com/x/status/1285507987864346630