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.

 

[Heartbreaker]

I was not aware they had configured it that way. But that sounds like a giant bottleneck if the GPU can only access half its RAM chips at any given time because it has to context switch back and forth between them.

You are out of your area of expertise. You always have a full 384 bit data bus, and your address bus determines which chips are part of that bus, on any access. There is no slowdown for this.

It's no different than if you had 12 x 2GB chips. That same address bit would then access one half of the chip or the other. You are never accessing everything simultaneously. The Address bus, chooses which chips and which portion of which chip is accessed.

The only reason they have 24 chips, is that 2GB GDDR6X chips aren't available yet. When they are, they will probably switch to a simpler 12 Chip cards for 24GB, but there really wont' be a performance difference.

Edit. I grew up in the 8bit computer era, and back in those days hacker/hobbyists would stack DIP memory chips together by piggyback soldering them together, with a leg or two free to change the addressing of the the new chip. This is one RAM chip directly soldered to another:




To the computer there is no difference if you have a single chip there or two at half the capacity, it's logically the same and performs the same. Here is another example, memory chips soldered directly to each other, with a an extra address line added in on an Apple II to increase memory capacity for cheaper:

Logically the 24 Chip 3090 memory is much the same as this. Each of of the 12 chips is directly paired with another on the reverse side of the board. There is no logical or performance difference than if they were single 2GB chips.

 

[trinibwoy]

I was not aware they had configured it that way. But that sounds like a giant bottleneck if the GPU can only access half its RAM chips at any given time because it has to context switch back and forth between them.

This isn’t a new technique and there’s no bottleneck. You’re still using the full bus width no matter which memory chips you read from. It’s the same approach used to make 48GB Quadros and Titans for several years now.

 

[MrTeal]

Still incorrect. These are are still 32 bit data bus chips on the 3090.

GA102 has 384 bit data bus. organized into 12 x 32 channels. It uses 10 of these channels on the 3080, for 10GB. It should be blindingly obvious, that It could use all 12 channels, for 12 GB on 3080 Ti.

On the 3090, it still uses 12x32 bit data bus. It just has two 32 bit, 1-GB chips on each channel, instead of one. The address bus determines which of two chips, on each channel, will respond.

Not sure how it works with GDDR6X, but with GDDR6 it would run in clamshell as well so 16 bits would run to one chip and 16 to the other, and so you'd need to fully populate all chips on the 3090 even if 2GB chips become available. For the 3080, all 32b would run to one chip.

 

[jim1976]

This is getting hilarious.
That's exactly what I suggested, then you went about to write a page long post about how memory bandwidth is not the same thing as memory capacity. Like... what the hell?

OK now I understand everything You're just making and changing goalposts as the conversation goes. I almost forgot which topic we were in, thanks for bringing me back to reality

Yes it did get hilarious and you absolutely did what you accuse me of.

Well, if their target audience were people who actually own a 4k display, that'd be a very-very sad market prognosis for the 3080. So with all due respect, no.

That's really a "wait to see" approach like the one you are trying to preach me about right?

Look you can lie to yourself all you want but don't try to sound objective because obviously you are not.. What suits your argument is on a need to wait validation basis from independent reviewers while all the rest are "preaching" from my part.. Let's just leave it here shall we? Your pretendious stance is making me laugh way more than I should.. Go quote someone else if what I write is bugging your "objective" stance lol

 

[Ajay]

You are out of your area of expertise. You always have a full 384 bit data bus, and your address bus determines which chips are part of that bus, on any access. There is no slowdown for this.

It's no different than if you had 12 x 2GB chips. That same address bit would then access one half of the chip or the other. You are never accessing everything simultaneously. The Address bus, chooses which chips and which portion of which chip is accessed.

The only reason they have 24 chips, is that 2GB GDDR6X chips aren't available yet. When they are, they will probably switch to a simpler 12 Chip cards for 24GB, but there really wont' be a performance difference.

Edit. I grew up in the 8bit computer era, and back in those days hacker/hobbyists would stack DIP memory chips together by piggyback soldering them together, with a leg or two free to change the addressing of the the new chip. This is one RAM chip directly soldered to another:

View attachment 32822


To the computer there is no difference if you have a single chip there or two at half the capacity, it's logically the same and performs the same. Here is another example, memory chips soldered directly to each other, with a an extra address line added in on an Apple II to increase memory capacity for cheaper:

Logically the 24 Chip 3090 memory is much the same as this. Each of of the 12 chips is directly paired with another on the reverse side of the board. There is no logical or performance difference than if they were single 2GB chips.

Wow, forgot about that. Didn't do this myself, but I remember buying workstation memory that came double stacked. I was kinda, wth? when it came in. Now we have dated ourselves!

 

[Heartbreaker]

Wow, forgot about that. Didn't do this myself, but I remember buying workstation memory that came double stacked. I was kinda, wth? when it came in. Now we have dated ourselves!

Yeah, commercial products also did this. I think I had memory expansion board for the Amiga 1000 that did this...

I think it added a whole 1MB of memory.

 

[maddie]

You are out of your area of expertise. You always have a full 384 bit data bus, and your address bus determines which chips are part of that bus, on any access. There is no slowdown for this.

It's no different than if you had 12 x 2GB chips. That same address bit would then access one half of the chip or the other. You are never accessing everything simultaneously. The Address bus, chooses which chips and which portion of which chip is accessed.

The only reason they have 24 chips, is that 2GB GDDR6X chips aren't available yet. When they are, they will probably switch to a simpler 12 Chip cards for 24GB, but there really wont' be a performance difference.

Edit. I grew up in the 8bit computer era, and back in those days hacker/hobbyists would stack DIP memory chips together by piggyback soldering them together, with a leg or two free to change the addressing of the the new chip. This is one RAM chip directly soldered to another:

View attachment 32822


To the computer there is no difference if you have a single chip there or two at half the capacity, it's logically the same and performs the same. Here is another example, memory chips soldered directly to each other, with a an extra address line added in on an Apple II to increase memory capacity for cheaper:

Logically the 24 Chip 3090 memory is much the same as this. Each of of the 12 chips is directly paired with another on the reverse side of the board. There is no logical or performance difference than if they were single 2GB chips.

I did this to the Atari ST 512 to turn it into a 1024 (KB). Good memories.

 

[lobz]

Yes it did get hilarious and you absolutely did what you accuse me of.



That's really a "wait to see" approach like the one you are trying to preach me about right?

Look you can lie to yourself all you want but don't try to sound objective because obviously you are not.. What suits your argument is on a need to wait validation basis from independent reviewers while all the rest are "preaching" from my part.. Let's just leave it here shall we? Your pretendious stance is making me laugh way more than I should.. Go quote someone else if what I write is bugging your "objective" stance lol

I'm sorry, man. I can't teach you to read and comprehend.

 

[Glo.]

https://twitter.com/x/status/1323994099012173824

Exciting stuff .

3050 Ti - 28 SMs, 3060 - 30 SMs.

 

[Stuka87]

https://twitter.com/x/status/1323994099012173824

Exciting stuff .

3050 Ti - 28 SMs, 3060 - 30 SMs.

And people say the 6800XT and 6900 are too similar...

 

[Glo.]

And people say the 6800XT and 6900 are too similar...

Well, one can argue, that because of this SM similarity, 3050 Ti will have 128 bit bus, instead of full 192 bit.

 

[beginner99]

Well, one can argue, that because of this SM similarity, 3050 Ti will have 128 bit bus, instead of full 192 bit.

and 4GB or RAM in 2020/21...like my 290x from 2013...

 

[gdansk]

One thing that's interesting about Ampere is that even the RTX 3060 will apparently be around 13 TFLOPs. Not that this translates to gaming performance but at raw compute that's basically a 2080 Ti. And in many CUDA benchmarks the 3070 is quite ahead of the 2080 Ti.

A good generation for that type of work, at least.

 

[Glo.]

https://twitter.com/x/status/1324035256341700608

No improvement in availability of Nvidia Ampere until March next year. From a very reliable source... I'm very sorry folks.

Crap.

 

[Hitman928]

https://twitter.com/x/status/1324035256341700608

Crap.

That would be shocking to me if true.

 

[Mopetar]

Is there some issue at Samsung we're not aware of because it isn't as though NVidia has to fight a whole bunch of other companies for limited wafers.

I know that earlier people had speculated that NVidia wouldn't launch until 2021, and while that turned out to be wrong, it's pretty obvious that NVidia launched as soon as they possibly could.

Still, it's unusual that the supply situation won't improve for another four months.

 

[Heartbreaker]

That would be shocking to me if true.

Demand should level off with AMD cards coming out, which for the most part look to have edged out NVidia this time.

I am interested in what the AMD 6600 and 6700 will look like, as even $500 is too much for me to spend on a GPU.

 

[gorobei]

the rumors were that the large ampere die size on samsung's 8nm(but really 10nm) is making for terrible yields. when you cant get that many good dies per wafer the production rate suffers. also if samsung did offer a pay per good die contract to nv(vs a typical pay per wafer), then they have no incentive to crank up production until they figure out the process issues.

 

[jpiniero]

the rumors were that the large ampere die size on samsung's 8nm(but really 10nm) is making for terrible yields. when you cant get that many good dies per wafer the production rate suffers. also if samsung did offer a pay per good die contract to nv(vs a typical pay per wafer), then they have no incentive to crank up production until they figure out the process issues.

8 nm is not exactly a new process. Albeit Samsung has never fabbed anything of GA102's size on the node.

 

[maddie]

the rumors were that the large ampere die size on samsung's 8nm(but really 10nm) is making for terrible yields. when you cant get that many good dies per wafer the production rate suffers. also if samsung did offer a pay per good die contract to nv(vs a typical pay per wafer), then they have no incentive to crank up production until they figure out the process issues.

If Nvidia launched anyway, in spite of production problems, can we say it's a launch or a reveal of coming attractions?

 

[Mopetar]

It's still a launch even if it's kind of a crummy one. There are people who have the cards in hand and you can go buy one yourself if you can find it in stock.

I also don't think AMD releasing cards will help much. There are a lot of dyed in the wool NVidia fans that won't get anything else and if the releases info about preorders taken vs. fulfilled are accurate and generalizable the there are a lot of people who'd need to cancel their order for the supply/demand to balance out.

I recall reading somewhere that the node NVidia is using has been customized for them. To what degree that's true or how much of a customization it represents I'm not sure of, but perhaps that could explain the shortages if Samsung is still in the process of bringing more of that online themselves.

 

[gorobei]

If Nvidia launched anyway, in spite of production problems, can we say it's a launch or a reveal of coming attractions?

typically everyone schedules 6 months of production time to build up stock for a new release. but if you have issues like intel pushing the schedule back 6 mo, you claim there wont be a delay by releasing without the 6 mo of stockpile time. no one can buy it but legally you released on time so stock owners cant complain.
nv in this case likely knew they wouldnt get volume out until 2021. but with big navi performing strong and the consoles hitting in nov, they pushed things up to stay relevant and grab some overpriced sales.

 

[mohit9206]

and 4GB or RAM in 2020/21...like my 290x from 2013...

Hopefully 6GB for 3050Ti as well but if it is 128 bit then unfortunately 4GB. Highly unacceptable considering it will almost assuredly cost more than 1650Ti(seems like a 1660S replacement).

 

[Glo.]

Is there some issue at Samsung we're not aware of because it isn't as though NVidia has to fight a whole bunch of other companies for limited wafers.

I know that earlier people had speculated that NVidia wouldn't launch until 2021, and while that turned out to be wrong, it's pretty obvious that NVidia launched as soon as they possibly could.

Still, it's unusual that the supply situation won't improve for another four months.

Its not an issue with Samsung process, but Ampere was not supposed to be released this year.

Nvidia simply rushed their release by few months, hence why we see the GPUs but there is not enough inventory. And its not due to absurdly high demand, but because of the actual lack of inventory.

 

[Glo.]

Hopefully 6GB for 3050Ti as well but if it is 128 bit then unfortunately 4GB. Highly unacceptable considering it will almost assuredly cost more than 1650Ti(seems like a 1660S replacement).

IMO, 3050 Ti will not cost more than 199$ MSRP. Regardless of VRAM buffer.