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.

 

[Det0x]

I guess we will have to wait for actual reviews then. The A100 PCIe has it's TDP dropped to 250 W yet nVidia didn't lower the quoted specs. Maybe it can boost well above the advertised frequencies and the PCIe version just can just not hold it as long, hence why nVidia says it's 10% slower.

The A100 PCIe has a TDP of 250W. According to ComputeBase who quotes NVIDIA, the card does indeed have lower TDP specs. For comparison, the SXM variant has a TDP of 400W. However, despite the lower TDP of the PCIe model, NVIDIA says that the peak power for both models is the same, only during a sustained load will provide 10 to 50% lower performance than SXM4 based variant.

 

[Stuka87]

Samsung's 10 nm process will give you worse Transistor density than even botched job that Navi 10 is on 7 nm TSMC.

TSMC's N7 is capable of packing 60 mln xTors/mm2 at high clock speeds. Navi 10 has 40 mln xTors/mm2.

I would expect at best 32-36 mln xTors/mm2 on Samsung's 10 nm node, which is called 8 nm. Its simply not that more dense than 12 nm FFN.

I would not call it "botched". I think AMD made it less dense on purpose in order to increase yields. Its not like they messed up and went "oops, we aren't very dense it seems". There is also something to be said to the better cooling characteristics of a less dense chip. And who knows, TSMC may even charge more when the chip density goes up.

 

[Ajay]

They can't if they want to keep their die sizes sane. Even 5376 ALU GPU will be over 600 mm2 with all the RT bells and whistles.

Never underestimate Nvidia's willingness to throw silicon at a problem to maintain leadership.

 

[Glo.]

Never underestimate Nvidia's willingness to throw silicon at a problem to maintain leadership.

Im not underestimating it. But that would require vast redesign. And as far as we have seen the rumors, the designs are set in stone. 84 SMs for 102 die. So the only thing that can be tweaked right now are the core clocks.

 

[DiogoDX]

The A100 PCIe has a TDP of 250W. According to ComputeBase who quotes NVIDIA, the card does indeed have lower TDP specs. For comparison, the SXM variant has a TDP of 400W. However, despite the lower TDP of the PCIe model, NVIDIA says that the peak power for both models is the same, only during a sustained load will provide 10 to 50% lower performance than SXM4 based variant.

Up to 50% throttling. So the peak numbers are just marketing.

 

[GodisanAtheist]

And that's a very good strategy by nvidia.
I think it's the correct thing to do with gpus. In the past ATI with the radeon 9700 did the same and it worked very well for them.

- I think the original statement is a bit disingenuous.

AMD and NV have been neck and neck on nodes for a LONG time, and there certainly has never been a node gap like we've have now since maybe the HD 3870 vs the GTX 8800 series (55nm vs 80/65 nm).

In that situation ATI/AMD was getting thrashed worse than they are now, and the node advantage really helped AMD get back into a competitive position with the follow-up 4800 series.

AMD might have beaten NV to the punch by a couple months after that on node transitions (40nm, 28nm come to mind) but NV actually released their 16nm cards before AMD released Polaris by a couple months.

 

[trexfromouterspace]

The A100 PCIe has a TDP of 250W. According to ComputeBase who quotes NVIDIA, the card does indeed have lower TDP specs. For comparison, the SXM variant has a TDP of 400W. However, despite the lower TDP of the PCIe model, NVIDIA says that the peak power for both models is the same, only during a sustained load will provide 10 to 50% lower performance than SXM4 based variant.

A simplified power comparison would indicate around 14.5% lower performance going from 400W to 250W. (250/400) ^ (1/3) = .855

Obviously not perfect since it assumes a perfectly linear 1:1 Voltage-Frequency curve and constant IPC, but enough for a general estimate. I would expect the impact to be closer to 10% than 50% for the vast majority of workloads.

 

[Ajay]

Im not underestimating it. But that would require vast redesign. And as far as we have seen the rumors, the designs are set in stone. 84 SMs for 102 die. So the only thing that can be tweaked right now are the core clocks.

Ah, didn't know that the number of SMs had been confirmed. Geez, not being on 7nm EUV really puts them in a bind.

 

[DXDiag]

Up to 50% throttling. So the peak numbers are just marketing.

NVIDIA says the 50% is rare, but it can happen in extreme situations. And no the peak numbers are just that: peak, same thing happened with V100 also, it had 4 different TDPs: 250w, 300w, 350w, 450w and even had a V100S variants with faster clocks and 250w TDP.

The NVIDIA numbers quoted for the A100 PCIe 250w are gainst the V100 PCIe 250w variant.

I guess we will have to wait for actual reviews then

Independent Systems integrated with A100 already confirm this:


"According to Brad Wheeler, the vice-president of the information technology branch of the university, 'next generation graphic cards from Nvidia' will be introduced. So we can expect that it will be about Ampere chips. According to Indiana University, these cards will be 70 to 75% faster than the existing models. "

Next generation of Nvidia graphics (Ampere/Hopper) up to 70% faster than Volta?

So here is the thing, the University of Indiana announced that it will delay the second phase of expansion of its super-computer Crey Shasta “Big Red II” to specifically wait for Nvidia's next ge...

www.guru3d.com

 

[Glo.]

Independent Systems integrated with A100 already confirm this:


"According to Brad Wheeler, the vice-president of the information technology branch of the university, 'next generation graphic cards from Nvidia' will be introduced. So we can expect that it will be about Ampere chips. According to Indiana University, these cards will be 70 to 75% faster than the existing models. "

Next generation of Nvidia graphics (Ampere/Hopper) up to 70% faster than Volta?

So here is the thing, the University of Indiana announced that it will delay the second phase of expansion of its super-computer Crey Shasta “Big Red II” to specifically wait for Nvidia's next ge...

www.guru3d.com

I'd like to hear Independent, second opinion on that. Best if it would be confirmed by actual benchmark numbers.

 

[Ajay]

NVIDIA says the 50% is rare, but it can happen in extreme situations. And no the peak numbers are just that: peak, same thing happened with V100 also, it had 4 different TDPs: 250w, 300w, 350w, 450w and even had a V100S variants with faster clocks and 250w TDP.

The NVIDIA numbers quoted for the A100 PCIe 250w are gainst the V100 PCIe 250w variant.

Independent Systems integrated with A100 already confirm this:


"According to Brad Wheeler, the vice-president of the information technology branch of the university, 'next generation graphic cards from Nvidia' will be introduced. So we can expect that it will be about Ampere chips. According to Indiana University, these cards will be 70 to 75% faster than the existing models. "

Next generation of Nvidia graphics (Ampere/Hopper) up to 70% faster than Volta?

So here is the thing, the University of Indiana announced that it will delay the second phase of expansion of its super-computer Crey Shasta “Big Red II” to specifically wait for Nvidia's next ge...

www.guru3d.com

That boost must be on some ML workload (TC), as FP didn't go up that much.

 

[sontin]

Sustained performance with one or two cards is 90% of the SMX4 version. Working more cards on a computation performance can drop to 50% because of PCIe 4.0.

But because the dual-slot add-in card form factor is designed for lower TDP products, offering less room for cooling and typically less access to power as well, the PCIe version of the A100 does have to ratchet down its TDP from 400W to 250W. That’s a sizable 38% reduction in power consumption, and as a result the PCIe A100 isn’t going to be able to match the sustained performance figures of its SXM4 counterpart – that’s the advantage of going with a form factor with higher power and cooling budgets. All told, the PCIe version of the A100 should deliver about 90% of the performance of the SXM4 version on single-GPU workloads, which for such a big drop in TDP, is not a bad trade-off.

NVIDIA Announces PCIe A100 Accelerator: 250 Watt Ampere In A Standard Form Factor

www.anandtech.com

 

[DXDiag]

That boost must be on some ML workload (TC), as FP didn't go up that much.

FP DID go up that much, A100 increased cache sizes significantly, helping to boost FP performance greatly. ML performance goes far beyond a mere 70%.

 

[RetroZombie]

AMD and NV have been neck and neck on nodes for a LONG time, and there certainly has never been a node gap like we've have now since maybe the HD 3870 vs the GTX 8800 series (55nm vs 80/65 nm).

And that line up also prove the same about the radeon 9700. A better refined design on an old process (like the GTX8800), beat an new rushed design on a new process (HD3870).

In that situation ATI/AMD was getting thrashed worse than they are now, and the node advantage really helped AMD get back into a competitive position with the follow-up 4800 series.

See only on the second refined design (the HD 4800) on the same node fixed the problems.
A bad comparison is the rerelease of vega 7nm, almost did nothing, still couldn't compete or matchup. The many months advantage of being on a new node was not threat to nvidia new designs on a old process.

AMD might have beaten NV to the punch by a couple months after that on node transitions (40nm, 28nm come to mind) but NV actually released their 16nm cards before AMD released Polaris by a couple months.

That time was GF vs TSMC, not TSMC vs TSMC so doesn't count, at least for me.

 

[Ajay]

FP DID go up that much, A100 increased cache sizes significantly, helping to boost FP performance greatly. ML performance goes far beyond a mere 70%.

View attachment 23885

In which case, why did Cray Shasta “Big Red II" only see a 70-75% increase 'on current models'? What are those models? Are they CPU/RAM/IO bound?

 

[GodisanAtheist]

And that line up also prove the same about the radeon 9700. A better refined design on an old process (like the GTX8800), beat an new rushed design on a new process (HD3870).

- I don't want to chase this point down the bottomless rabbit hole it can be, because I don't really agree with the fundamental point that there is some sort of connection between process node and good cards, but the HD 2900 XT was the original contender against the GTX 8800, and both were on the 80nm process. The 2900XT was such an unmitigated disaster (despite being on an "old process") that AMD had to basically shrink the thing and throw out its memory bus before it began behaving like a card in its performance class should.

 

[RetroZombie]

I don't really agree with the fundamental point that there is some sort of connection between process node and good cards

Neither do I, my point was risking an bigger chip on old process is a 'better chip' than trying to do an smaller chip on new process.
By better, I mean time to market is faster, better yields, get more time to refine and tune it since the process is already better known, ... overall the end up design should be faster.

 

[GodisanAtheist]

Neither do I, my point was risking an bigger chip on old process is a 'better chip' than trying to do an smaller chip on new process.
By better, I mean time to market is faster, better yields, get more time to refine and tune it since the process is already better known, ... overall the end up design should be faster.

- Fair enough. Intel (and Nvidia's) tick-tock strategy was a solid play for as long as they can make it work.

AMD has started to implement something like this on the CPU side, but is still lacking on the GPU front. Just never had the resources to stagger both a node shrink team and an arch team it seems.

 

[Konan]

Interesting! Total rumour territory here, doesn't quite add up in places with other leaks/rumours.

https://twitter.com/x/status/1275331994893684736

Rumor: Alleged Spec Sheet Of An NVIDIA RTX 3080 Ti Leaked, Shows Up To 23 TFLOPs Of Peak Graphics Horsepower

A screenshot of an alleged spec sheet of the NVIDIA RTX 3080 Ti has leaked out via Twitter (@CyberPunkCat). I will say upfront that the leak has some oddities, technical inconsistencies and is not verified and is therefore in solid rumor territory. As with all rumors - a grain of salt is highly...

wccftech.com

 

[Veradun]

Interesting! Total rumour territory here, doesn't quite add up in places with other leaks/rumours.

https://twitter.com/x/status/1275331994893684736

Rumor: Alleged Spec Sheet Of An NVIDIA RTX 3080 Ti Leaked, Shows Up To 23 TFLOPs Of Peak Graphics Horsepower

A screenshot of an alleged spec sheet of the NVIDIA RTX 3080 Ti has leaked out via Twitter (@CyberPunkCat). I will say upfront that the leak has some oddities, technical inconsistencies and is not verified and is therefore in solid rumor territory. As with all rumors - a grain of salt is highly...

wccftech.com

As fake ad you can get

 

[uzzi38]

We're seriously giving time of the day who made up the SK Hynix Big Navi rumour?

Come on guys. Come on.

 

[Konan]

We're seriously giving time of the day who made up the SK Hynix Big Navi rumour?

Come on guys. Come on.

Yeah, you’re right there. Didn’t connect that he was the same guy with the fake big Navi rumour.
Didn’t look right to me from the start.

 

[JoeRambo]

That spec might be completely fake, but that is where we should expect new cards to be at clocks/shaders/memory bw wise.

 

[uzzi38]

Yeah, it is around my expectations personally, I just strongly dislike giving people known to be making things up time of day.

You can take as a good a guess as you want, it's not OK when you try to play that guess off as something your "sources" tell you.

 

[Baasha]

Interesting! Total rumour territory here, doesn't quite add up in places with other leaks/rumours.

https://twitter.com/x/status/1275331994893684736

Rumor: Alleged Spec Sheet Of An NVIDIA RTX 3080 Ti Leaked, Shows Up To 23 TFLOPs Of Peak Graphics Horsepower

A screenshot of an alleged spec sheet of the NVIDIA RTX 3080 Ti has leaked out via Twitter (@CyberPunkCat). I will say upfront that the leak has some oddities, technical inconsistencies and is not verified and is therefore in solid rumor territory. As with all rumors - a grain of salt is highly...

wccftech.com

lel