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.

 

[Hitman928]

Gotta make room for a 3080 Ti later on.

Leaves room for some supers to slot in...

Yeah, but there's not much to squeeze in there. The 3090 only has 20% more cuda cores and is 350W versus 320W of the 3080. So a 3080 Ti/Super is suppose to come in with 10% more performance than a 3080 for $999? Or are they going to match the 3090 at $999? There's not really room for another card in between that makes sense unless they put out the 3090 as just a huge money grab and then release a 3080 super at the same performance for over 40% off just a little bit later.

 

[aleader]

Yep, if it turns out true that the 3070 is a 2080 Ti class GPU for $499 then I'd call that a huge win. Will sell like hotcakes.

Yes, 3070 it (likely) is for me. It will last me 10 years at 1440p. Plenty of juice for $499. I will still wait for RDNA 2 to see what they've got. This should finally bring down the ridiculous prices for the old cards now too, unless people really do have no common sense. I see Newegg prices still haven't budged.

 

[ScottAD]

Watching the GPU resale market plummet r/n...Reddit HWS is hilarious.

 

[TESKATLIPOKA]

Again, no one will ever need so much power to play games, so whatever "Big" Navi delivers is enough. But as I knew too well there's no way for AMD to catch up with Nvidia. The ludicrous idea that AMD was on it's way to pass Nvidia within two generations was pure insanity.

If you mean Big Navi won't match RTX 3090 then I think so too, but for the rest I am not so sure, probably between RTX 3080 and RTX 3070.
2x FP32 was a pleasant and unexpected surprise, but just doubling of Cuda cores doesn't mean double e close to double performance If TMUs and ROPs weren't doubled.
Another question is why is RTX 3080 or RTX 3070 so cheap?

 

[sze5003]

A 3080Ti could be 1 year out, its unfortunately a gamble to wait for it, the whole situation could be different by then.

Weird why one year out and not couple more months? I do need a new card and I want something I can keep as long as I've had my 1080ti. Not sure I'll want to wait.

 

[dr1337]

Big marketing mistake to call the 3090 that. Calling it Titan would at least 'justified' its price in the eyes of many. Especially after he called the 3080 the 'flagship'.

They're using marketing to compete with AMD and to preemptively deny the x90 teir GPU. Its not about justifying the price but more about 'showing performance' compared to any, say 6900XT gpus.

 

[TESKATLIPOKA]

Isn't it the same as Dual CU in RDNA1 being able to execute two wavefronts?

And the number of ALUs is still the same?

So how many TMUs are there? Are they doubled too or not?

 

[moonbogg]

Definitely waiting on the 3080Ti. $700 is better than expected, but I can't downgrade on Vram while upgrading in other areas. A next-gen GPU at $700 shouldn't be a pseudo-upgrade in any way. I think Lisa is going to kick his ass. I'd expect a 3080Ti next year or something, possibly sooner depending on the competition. I'm looking forward to those 3090 AIB markups for the epic lulz though.

 

[lobz]

Seems like Kopite nailed everything down.
PCB, Process, G6X, Cores. Everything...

Only thing that stands out is that, why are they so affordable?

Last minute decision by Mr. Huang.

 

[Hi-Fi Man]

"2x FP32" is interesting. Would like to get some details from Ryan on what's going on there. Perhaps each CUDA core is a superscaler dual pipelined core now? If so, it's a bit disingenuous to just double the CUDA core numbers. I would be actually surprised if "2x FP32" means two separate CUDA cores that are somehow tied together.

 

[Glo.]

So how many TMUs are there? Are they doubled too or not?

I don't know.

Is it meaningful for ALU counts?

With this presentation Nvidia has not shown more than they did, and only raised a lot of questions, instead of answering them...

 

[Hitman928]

If you mean Big Navi won't match RTX 3090 then I think so too, but for the rest I am not so sure, probably between RTX 3080 and RTX 3070.
2x FP32 was a pleasant and unexpected surprise, but just doubling of Cuda cores doesn't mean double performance If TMUs and ROPs weren't doubled.
Another question is why is RTX 3080 or RTX 3070 so cheap?

Yeah, I'm pretty confident the 3090 is out of reach of big Navi and Nvidia is pricing it accordingly. The 3080 pricing puts pressure on AMD and is where I expect the real battle with big Navi to take place but obviously we'll have to see what AMD has first to be sure.

 

[Konan]

For reference:

 

[CakeMonster]

Weird why one year out and not couple more months? I do need a new card and I want something I can keep as long as I've had my 1080ti. Not sure I'll want to wait.

Well we just went 2 years with the 2080Ti as the flagship ~30% ahead of the 2080, and the Super cards arrived in between at the 1 year mark.

 

[StinkyPinky]

Big Navi doesn't need to compete with the 3090. That's a tiny market. Now the 3070s and 3080s, it does need to compete.

 

[TESKATLIPOKA]

I don't know.

Is it meaningful for ALU counts?

With this presentation Nvidia has not shown more than they did, and only raised a lot of questions, instead of answering them...

For ALU count not but for performance I think yes.

 

[lobz]

Now we are trying hard, aren't we?

What are you talking about, we genuinely don't know and trying to make conversation, what the effect of doubling the FP ALUs will have and how it resembles to other architectures. Please stop being so childish, we're all happy for the announced low prices and it's only you who can't get over getting trumped by your own arrogance when you were bashing everyone who didn't believe you that it was going to be 7nm. Well, it's not 7nm, get over it.

 

[lobz]

In a weird sort of turnabout, AMD now has the perf per Flop advantage.

Jokes aside, Pricing for the 3070 and 30870 is better than expected. I'm impressed by this launch, but it's clear that that's for a reason

Dude that was hilarious

 

[Insomniator]

Very little to complain about here from what I saw. Will be funny seeing people trying to sell used 1080TI's for $500 still.

3080 for $700 seems great if they are actually available. Same for the 3070.

3090 looks good as a halo, I'm sure people will cry about it but they can charge whatever the **** they want for the halo card (I might preorder one).

 

[n0x1ous]

new NV link bridge required? weird since no mention of SLI at all

 

[sze5003]

Very little to complain about here from what I saw. Will be funny seeing people trying to sell used 1080TI's for $500 still.

3080 for $700 seems great if they are actually available. Same for the 3070.

3090 looks good as a halo, I'm sure people will cry about it but they can charge whatever the **** they want for the halo card (I might preorder one).

I'm pretty sure you won't be able to get a 3080 for their set starting price. I may try and pre order a 3090 and see if it fits in my mid size atx case.

 

[Det0x]

Does anyone know when the review embargo ends ??

 

[Glo.]

The good news is: used market is going to be an interesting place right now .

 

[randomhero]

"2x FP32" is interesting. Would like to get some details from Ryan on what's going on there. Perhaps each CUDA core is a superscaler dual pipelined core now? If so, it's a bit disingenuous to just double the CUDA core numbers. I would be actually surprised if "2x FP32" means two separate CUDA cores that are somehow tied together.

I think that is it. Turing could do one int and one fp op per alu per clock. Ampere can probably do 2 fp or one fp and one int op per clock. Ultimately you are always limited by bandwidth. So you see around max 30% uplift in performance if I read that chart correctly(3080 above 2080ti).

 

[pj-]

RTX IO seems interesting. Does that erase the big advantage of PS5/XBSX over pc? Will big navi have something similar?