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.

 

[nurturedhate]

It is real simple, if you cannot properly cite your work you don't get credit.

Read the thread.

Anyone who is too lazy to read the thread they are posting in please put me on ignore. If I wanted to partake in sound bite stupidity I'd go the Twitter route.

Reading a thread and having a decent conversation isn't done by focusing on the person talking, but the merit of the idea and the perspective it brings, then counter perspective it offers from others.

You want a personal discussion, there are forums for that, this is a tech forum.

You refuse to properly cite your work. You purposely remove context from posts. You obfuscate.

 

[exquisitechar]

https://twitter.com/x/status/1288811296435998723

In line with my expectations.

 

[DXDiag]

In TimeSpy Extreme the 2080Ti FE scores around 6400, If the 3090 scores 10000 it means it is 56% faster in that benchmark.

NVIDIA GeForce RTX 2080 Ti Review: No Seriously, Just Buy It

NVIDIA's new GeForce RTX 2080 Ti is the best graphics card ever made, period.

www.tweaktown.com

 

[jpiniero]

https://twitter.com/x/status/1288811296435998723

In line with my expectations.

Guess there's a question as to whether the 3090 is just what they are calling the Titan.

 

[Bubbleawsome]

If the 3090 is Titan Class and the 3080/3080ti slots in about 15% under it, then a top-tier AMD part being 20% better than a 2080ti leaves it about 12% behind the top-tier consumer card. Not perfect, but not bad especially if they keep prices reasonable. I'd take a 12% performance loss if it's more in the $800 range than the $1200 range like the 2080ti

 

[Stuka87]

If the 3090 is Titan Class and the 3080/3080ti slots in about 15% under it, then a top-tier AMD part being 20% better than a 2080ti leaves it about 12% behind the top-tier consumer card. Not perfect, but not bad especially if they keep prices reasonable. I'd take a 12% performance loss if it's more in the $800 range than the $1200 range like the 2080ti

The Titan has not been a consumer card for some time now though. Very few people would spend $2500 on a card with same/worse gaming performance than the Ti model below it. Its price increase is for the compute abilities.

Its possible nVidia will bring its price back down to the 1200 range, but seems unlikely.

 

[BenSkywalker]

I think what he's saying, and I'm half expecting it, is that the Titan will be renamed to 3090 and it will take over the halo position for consumer cards. If the rumors about the 3090 having 24GB of RAM are true this would make sense.

If they do that, I'd expect the pricing to be adjusted down in terms of tiers. Maybe $700 80, $900 Ti, $1300 90?

Personally I'm expecting this gen to have a decent amount more ML hardware so it may be the Titan just wasn't making sense anymore.

I could be way off base, but that's what it looks like.

 

[MrTeal]

I wouldn't expect a $900 Ti, certainly not for the launch models. I would actually be pleasantly surprised if you could buy any 3080 Ti under $1000 day 1.

 

[BenSkywalker]

The Ti is going to be much smaller this generation. Between the backlash over the pricing for last generation and the hype surrounding the AMD launch I'd expect nVidia to be a bit more aggressive with pricing.

Also keep in mind that the current Ti has a msrp of $999- the inflated prices we are seeing now are due to supply limitations. That shouldn't be nearly as big of an issue with a much smaller die.

 

[Konan]

Also keep in mind that the current Ti has a msrp of $999- the inflated prices we are seeing now are due to supply limitations. That shouldn't be nearly as big of an issue with a much smaller die.

The prices have been shitty with Turing since they came out and with no supply issues several months post launch the price is still crap.
I feel the pricing will be better this time around but not by much. Nvidia can go first and then AMD will undercut without much performance difference and then things should equalize for a bit.

A 2080 Founders Edition @ $800 launch was a kick to the nuts. The standard model at $700 wasn't much better. We'd be lucky to see a 3080 "founders" for $700.

 

[Stuka87]

The prices have been shitty with Turing since they came out and with no supply issues several months post launch the price is still crap.
I feel the pricing will be better this time around but not by much. Nvidia can go first and then AMD will undercut without much performance difference and then things should equalize for a bit.

A 2080 Founders Edition @ $800 launch was a kick to the nuts. The standard model at $700 wasn't much better. We'd be lucky to see a 3080 "founders" for $700.

And then they came out with the 'Super' models, for that extra 1-2 punch.

 

[Tup3x]

I'd hope that they'd just stop using Ti and Super suffixes. For example the jump in performance between 2080 and 2080 Ti was so large that it would have made more sense to just call it 2090. Same thing for previous generations too. If they want to make a refresh then those suffixes or increasing the number by five for example would be fine. Initially just numbers would make more sense.

 

[jpiniero]

I'd hope that they'd just stop using Ti and Super suffixes. For example the jump in performance between 2080 and 2080 Ti was so large that it would have made more sense to just call it 2090. Same thing for previous generations too. If they want to make a refresh then those suffixes or increasing the number by five for example would be fine. Initially just numbers would make more sense.

The gap shouldn't be as large as 2080->2080 Ti; it's only ~20% more shaders if it ends up being the same counts as the rumors (4352 vs 5248/5376).

 

[Veradun]

I actually wish it was against the rules. Not only does it remove any context and make it harder to follow, but the person being quoted it not notified that they have been quoted, so they are unlikely to give a rebuttal.

There is an ignore option anyway

 

[jpiniero]

https://twitter.com/x/status/1288833401907290112

Looked at this again, does seem like this is the 3080. If you look at the MSI Lightning, seems that's a ~320 W card. Comparing the efficency doesn't look as bad if you compare the 3080 to the Lightning Z, esp if nVidia can get another 5-10% in clocks in the final version.

 

[uzzi38]

https://twitter.com/x/status/1288833401907290112

Looked at this again, does seem like this is the 3080. If you look at the MSI Lightning, seems that's a ~320 W card. Comparing the efficency doesn't look as bad if you compare the 3080 to the Lightning Z, esp if nVidia can get another 5-10% in clocks in the final version.

No, it's not. It's a higher tier GPU that's crippled by poor memory. The 12Gbps is reporting correctly.

 

[jpiniero]

No, it's not. It's a higher tier GPU that's crippled by poor memory. The 12Gbps is reporting correctly.

I'm not entirely sure how to calculate the bandwidth, but I have a feeling the 3080 will be more than 256 bit like usual. If it was 320 bit, that would be 15, right?

 

[Konan]

For what it is worth Kopite7kimi is proposing the TimeSpy Extreme (Graphics) scores:
3080 = almost 8600 (Just about 36% faster than a stock RTX 2080 Ti) (aftermarket models about ~20% faster, but we also don't know what a 3080 aftermarket model can do yet)
3090 = almost 10000 (about 50% faster than a stock RTX 2080 Ti)
And tweaks, QA, drivers etc not finished yet..

 

[jpiniero]

For what it is worth Kopite7kimi is proposing the TimeSpy Extreme (Graphics) scores:
3080 = almost 8600 (Just about 36% faster than a stock RTX 2080 Ti) (aftermarket models about ~20% faster, but we also don't know what a 3080 aftermarket model can do yet)

20% is presumably at the "official" TDP, which will be much less than the 320 W of the FE.

 

[uzzi38]

I'm not entirely sure how to calculate the bandwidth, but I have a feeling the 3080 will be more than 256 bit like usual. If it was 320 bit, that would be 15, right?

If it's using GA102, it's probably going to be a 384-bit bus.

 

[BenSkywalker]

If it's using GA102 then how many parts are there- ie is there going to be a 3090/Titan, 3080Ti and a 3080 or will there just be 3090/3080?

If it's three tier we could see 384/352/320 or perhaps 384/320 but with different speeds?

Some of the rumors are saying the 70Ti will be coming back too, so in theory we could be looking at five parts above the x60.. Not saying won't happen, but if it does that would lead to either some very granular price increments or stone outrageous prices inn the high end.

 

[Veradun]

If it's using GA102, it's probably going to be a 384-bit bus.

They'll probably also cut the bus width on cut down versions like the 3080 seems to be. 320b seems a good guess to me :>

 

[BFG10K]

AMD's now shipping 16GB 5700XTs for iMacs: https://www.techpowerup.com/270653/...ics-performance-to-updated-27-inch-apple-imac

If nVidia only ships 10-12GB on high-end Ampere, that's utter garbage.

 

[Mopetar]

I would be incredibly surprised if NVidia only offered at most 12 GB in high end Ampere.

I think that would be more surprising than just about anything else that could happen with it, even given some of the other rumors that seem highly unlikely.

 

[Bouowmx]

2 GB GDDR6 chips appear to nerf bandwidth on the Radeon Pro RX 5700: 384 GB/s (12 GT/s) instead of 448 GB/s (14 GT/s) on the regular versions.
Though NVIDIA Turing isn't affected even on the Quadro RTX 8000 48 GB, with 2 GB and double-sided memory: still 14 GT/s vs GeForce RTX 2080 Ti.

I guess it will be 1 GB chips for Ampere (12 GB on 384-bit) to get higher bandwidth: the alleged >18 GT/s for GeForce RTX 3080 and 3080 Ti/3090.