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.

 

[Markfw]

Yea well then you used them to get money back so then you didn't completely lose out. As long as you use them for something worth your while, it's not a lost cause. I skipped 20 series all together after reviews were out and I'm glad I did because I didn't do much more than what I needed with my 1080ti. I'll probably bite on the 3090 now though too. Tired of doing the yearly upgrade.

I used then for F@H, first to cure cancer, than Covid crap. No actual monetary payback.

 

[Stuka87]

More details on the 3000 series coolers:

 

[traderjay]

ROG Strix GeForce RTX 3090 24GB GDDR6X | Graphics Cards

The ROG Strix GeForce RTX 3090 24GB GDDR6X unleash the maximum performance on NVIDIA Ampere Architecture, by featuring Axial-tech fan design, 0dB technology, 2.9-slot Design, Dual BIOS, Auto-Extreme Technology, SAP II, MaxContact, and more.

www.asus.com

My case takes 32 cm GPU length...this card is 31.85cm... Whew...can't wait

 

[A///]

ROG Strix GeForce RTX 3090 24GB GDDR6X | Graphics Cards

The ROG Strix GeForce RTX 3090 24GB GDDR6X unleash the maximum performance on NVIDIA Ampere Architecture, by featuring Axial-tech fan design, 0dB technology, 2.9-slot Design, Dual BIOS, Auto-Extreme Technology, SAP II, MaxContact, and more.

www.asus.com

My case takes 32 cm GPU length...this card is 31.85cm... Whew...can't wait

Anything can fit if you take a cutting wheel to it!

 

[Stuka87]

ROG Strix GeForce RTX 3090 24GB GDDR6X | Graphics Cards

The ROG Strix GeForce RTX 3090 24GB GDDR6X unleash the maximum performance on NVIDIA Ampere Architecture, by featuring Axial-tech fan design, 0dB technology, 2.9-slot Design, Dual BIOS, Auto-Extreme Technology, SAP II, MaxContact, and more.

www.asus.com

My case takes 32 cm GPU length...this card is 31.85cm... Whew...can't wait

That's cutting it pretty close. I have learned to not always trust case max GPU lengths. Hope it fits without cutting things

EDIT: Fixed typo :/

 

[Midwayman]

I have learned to not always trust case mac GPU lengths.

Friggen apple.

 

[Stuka87]

Friggen apple.

Ahh typos

 

[GoodRevrnd]

More details on the 3000 series coolers:

Just watched this earlier, certainly timely. I'm thinking through my next build and was wondering about the impact on that top vent. Currently thinking about an NR200 and what the best config would be. GPU pulling through bottom sounds good if off the floor. AIO out the side, 2 front intakes.

 

[Stuka87]

Just watched this earlier, certainly timely. I'm thinking through my next build and was wondering about the impact on that top vent. Currently thinking about an NR200 and what the best config would be. GPU pulling through bottom sounds good if off the floor. AIO out the side, 2 front intakes.

With a 3 slot GPU (3080 and 3090), you can't use the bottom fans in the NR200. And I am betting the 3090 is too long for that case. For some reason Cooler Master doesn't list max GPU length for that case.

 

[mohit9206]

If Nvidia has a 3070Ti ready to go for $599 with 16GB to counter Navi, I don't understand how it could be a good option. It will basically be a 3070 with slightly higher clock speeds and additional 8gb ram. But for $100 more you could have a much faster card with less ram. It would be kind of similar situation to GTX 780 6GB and 780Ti 3GB.

 

[Kenmitch]

Just watched this earlier, certainly timely. I'm thinking through my next build and was wondering about the impact on that top vent. Currently thinking about an NR200 and what the best config would be. GPU pulling through bottom sounds good if off the floor. AIO out the side, 2 front intakes.

Looked interesting, but he sure did babble on forever with how he's going to test it. Never the less it should be worth watching in the end.

 

[Mopetar]

If Nvidia has a 3070Ti ready to go for $599 with 16GB to counter Navi, I don't understand how it could be a good option. It will basically be a 3070 with slightly higher clock speeds and additional 8gb ram. But for $100 more you could have a much faster card with less ram. It would be kind of similar situation to GTX 780 6GB and 780Ti 3GB.

That just seems like a confusing product stack. Unless AMD really throws NVidia a curveball I don't see a 3070 Ti coming out soon.

Instead I expect we see the same thing as with Turing where there's an eventual Super line across the various models where we get better dies with that have fewer parts disabled.

The 3070 could always get a bump from going to GDDR6X. Depending on when that launches prices might allow for a 16 GB version. Or maybe they'll find someone to make 1.5 GB memory chips and we'll get a 12 GB model which is probably more than enough for what the card would need to be.

 

[Ajay]

The 3070 could always get a bump from going to GDDR6X. Depending on when that launches prices might allow for a 16 GB version. Or maybe they'll find someone to make 1.5 GB memory chips and we'll get a 12 GB model which is probably more than enough for what the card would need to be.

Micron has 2 GB chips in the pipeline. They are due sometime next year and would probably make a nice refresh (assuming the unit prices say within the limits of the targeted addressable market).

 

[FatherMurphy]

Great info and post.
I'd like to add the caveat that A100 has 40 MB of L2 cache, which likely inflated the transistor density. Navi 10 is on TSMC N7P and it only gets around 41 MT/mm2 (10.3B xtors, 251mm2) so for a typical gaming GPU, TSMC N7 and SS 8N are comparable. We'll see how much L2 cache GA102 had but I wouldn't be surprised if it's far less than the 40MB of A100.

Your point is well made. The A100 has 40MB of L2 cache, but GA102 has only 6MB (which seems really low) for a full chip (the 3080 has 5MB enabled). https://wccftech.com/nvidia-geforce-rtx-30-series-ampere-graphics-cards-deep-dive/

 

[T4C Fantasy]

i have added a nice list here comparing densities

Nvidia Graphics IP

Graphics ChipDirectXShader ModelWDDMOpenGLOpenCLVulkanCUDAPureVideoVDPAUFP16FP64 NV11.0N/AN/AN/AN/AN/AN/AN/AN/AN/AN/A NV35.0N/AN/A1.0N/AN/AN/AN/AN/AN/AN/A NV45.0N/AN/A1.2N/AN/AN/AN/AN/AN/AN/A NV56.0N/AN/A1.2N/AN/AN/AN/AN/AN/AN/A NV10/11/157.0N/AN/A1.2N/AN/AN/AN/AN/AN/AN/A NV17/18/198.01.1 /...

www.techpowerup.com

 

[GoodRevrnd]

With a 3 slot GPU (3080 and 3090), you can't use the bottom fans in the NR200. And I am betting the 3090 is too long for that case. For some reason Cooler Master doesn't list max GPU length for that case.

It supports 330mm length. With how this case is laid out I was thinking the GPU fans act as the bottom intakes since the card is positioned so low and the bottom panel is relatively open.

 

[Panino Manino]

Was already discussed that these "doubled" cores may not have be really "doubled"? For what I understood Nvidia took the core and "extended" instead of making it two. Would be a "Constructor Core" moment? Just one core that sometimes with proper optimization can do the work of two?

 

[Konan]

Was already discussed that these "doubled" cores may not have be really "doubled"? For what I understood Nvidia took the core and "extended" instead of making it two. Would be a "Constructor Core" moment? Just one core that sometimes with proper optimization can do the work of two?

Yes they have doubled. Take a look through Ryan's tweets.

https://twitter.com/x/status/1301990126852034561

 

[dr1337]

For what I understood Nvidia took the core and "extended" instead of making it two.

Yeah this is pretty much what they did. They can do either 1 int32 + 1 fp32 or 2x fp32 per clock, meaning that core utilization will never be at 100% (tbf this wouldnt matter if you're only running fp32 but the int silicon is still there not being used).

 

[maddie]

Yeah this is pretty much what they did. They can do either 1 int32 + 1 fp32 or 2x fp32 per clock, meaning that core utilization will never be at 100% (tbf this wouldnt matter if you're only running fp32 but the int silicon is still there not being used).

They are not the same shader cores. If you want, you can say they "extended" the int cores to also be used as fp cores and arranged these new cores to be 1/2 the total cores. There would be common circuitry to both an int or fp operation so this is a good way to improve processing with small cost in area. Nice work by Nvidia here.

Turing had a 1:1 fp:int core ratio and any software not exactly that ratio had idle cores with either the fp or int cores as the limiting path. It appears that they do not see int operations as being more than 50% of the total, but it can be less. Ampere can go from 1:1 to 1:0 fp:int ratio and still use all the cores to speed up processing.

 

[moonbogg]

I watched an AdoredTV video where, as long as I understood correctly, that the CUDA cores are physically half of what is being shown and they are counted as double because of some hyperthreading-type of technique being used. I also saw another DOOM Eternal 4K video between an overclocked 2080Ti vs the 3080, and this time the 3080 was only about 10% faster. Not sure if that 3080 can be overclocked to restore the delta. He also raised the question of what the delta would be at lower resolutions. Perhaps it would be somewhat less? I think putting my F5 finger on ice and reading reviews this time is called for. I knew what I was getting with the 1080Ti, so that was an easy choice. This has a slight Turing feeling to it where you don't know what performance really is at the end of the day in YOUR use case, how well they OC etc.

 

[IntelUser2000]

Remember the obsession with clock speeds? It mattered because until then, faster clockspeeds meant higher performance.

Maybe they are doing the same with compute units lol. Less technical people will say the new Geforce has 10K cuda cores.

 

[AtenRa]

I watched an AdoredTV video where, as long as I understood correctly, that the CUDA cores are physically half of what is being shown and they are counted as double because of some hyperthreading-type of technique being used. I also saw another DOOM Eternal 4K video between an overclocked 2080Ti vs the 3080, and this time the 3080 was only about 10% faster. Not sure if that 3080 can be overclocked to restore the delta. He also raised the question of what the delta would be at lower resolutions. Perhaps it would be somewhat less? I think putting my F5 finger on ice and reading reviews this time is called for. I knew what I was getting with the 1080Ti, so that was an easy choice. This has a slight Turing feeling to it where you don't know what performance really is at the end of the day in YOUR use case, how well they OC etc.

That is incorrect , there are double the FP32 ALUs

 

[AtenRa]

Remember the obsession with clock speeds? It mattered because until then, faster clockspeeds meant higher performance.

Maybe they are doing the same with compute units lol. Less technical people will say the new Geforce has 10K cuda cores.

It has , its dual 16x FP32 ALUs

 

[IntelUser2000]

It has , its dual 16x FP32 ALUs

I didn't say it didn't. Neither did I claim it previously.

They increased compute cores without boosting anything else so the gain isn't as big as expected.