Question 'Ampere'/Next-gen gaming uarch speculation thread

Page 85 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

Ottonomous

Senior member
May 15, 2014
559
292
136
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.
 

A///

Diamond Member
Feb 24, 2017
4,351
3,158
136
The 3090 appears to be replacing the Titan, which currently sells for $3000. So you could look at it as a 33% price cut :p

And why would, if the rumor is true, NVidia cut 33% from the Titan's price if the 3090 is replacing it unless a Titan may come out later on? Is this new segmentation or reshuffling names? If the 3090 matches last gen Titan performance or exceeds it to a considerable degree, is this NVidia cementing their fear that AMD have actually come up with something viable?

*Insert UFO sounds*
 

FaaR

Golden Member
Dec 28, 2007
1,056
412
136
If those GPUs are having 24 memory chips its means 120W in power draw from memory subsystem, alone.
Not necessarily. Those memory chips would be running in GDDR clamshell mode where each pair of chips have half their memory bus interface disabled, so power would likely not increase (as) much as raising chip count from 12 to 24 would otherwise suggest. As it's the PHY which consumes the vast majority of power in a GDDR device, if you disable half of it, almost half of the chip's power dissipation goes away.

You'd raise consumption by whatever running some on-chip logic requires, to manage chip functions and executing commands sent to it (reading, writing and so on), and keeping the DRAM arrays refreshed, which would be pretty manageable.
 
  • Like
Reactions: Konan

maddie

Diamond Member
Jul 18, 2010
4,881
4,951
136
And why would, if the rumor is true, NVidia cut 33% from the Titan's price if the 3090 is replacing it unless a Titan may come out later on? Is this new segmentation or reshuffling names? If the 3090 matches last gen Titan performance or exceeds it to a considerable degree, is this NVidia cementing their fear that AMD have actually come up with something viable?

*Insert UFO sounds*
Pricing will be quite interesting as they're launching first.
 

Stuka87

Diamond Member
Dec 10, 2010
6,240
2,559
136
And why would, if the rumor is true, NVidia cut 33% from the Titan's price if the 3090 is replacing it unless a Titan may come out later on? Is this new segmentation or reshuffling names? If the 3090 matches last gen Titan performance or exceeds it to a considerable degree, is this NVidia cementing their fear that AMD have actually come up with something viable?

*Insert UFO sounds*

It will most likely lack some of the DP performance Titan has. RTX Titan has been pretty poor on sales. With its price being so high, people just go up to a Quadro. And nobody buys a Titan for its gaming performance. So release something in its place, gimp the workstation performance, and sell it as a top tier gaming card.
 

A///

Diamond Member
Feb 24, 2017
4,351
3,158
136
Pricing will be quite interesting as they're launching first.

Definitely. I made a post last night about old NVidia prices and how it's changed so much. I'm both excited and afraid of the launch. If NVidia get forced into a yearly game of launches by AMD I'm okay with that and would only bother upgrading every 2-3 years if the performance difference is decent enough to warrant the upgrade.
It will most likely lack some of the DP performance Titan has. RTX Titan has been pretty poor on sales. With its price being so high, people just go up to a Quadro. And nobody buys a Titan for its gaming performance. So release something in its place, gimp the workstation performance, and sell it as a top tier gaming card.
True. I know two people who buy Titans for gaming, but they have incredible rigs and incredibly deep wallets for their hobbies. I'm really interested in sales figures concerned regular GeForce vs. Quadro vs. Titan in professional fields since the go to advice on modeling sites seems to be save your money and go for GeForce. It used to be exclusively Quadro back in the day.
 
  • Like
Reactions: Konan

Thala

Golden Member
Nov 12, 2014
1,355
653
136
Interesting. It basically means that memory subsystem alone draws 60W per 12 memory chips.

If those GPUs are having 24 memory chips its means 120W in power draw from memory subsystem, alone.

o_O

Lol, power scales with bandwidth not with size...and GDDR6X has higher bandwidth efficiency.
Anyway your calculation is nonsense.
 
  • Like
Reactions: Lodix and DXDiag

Glo.

Diamond Member
Apr 25, 2015
5,803
4,777
136
Lol, power scales with bandwidth not with size...and GDDR6X has higher bandwidth efficiency.
Anyway your calculation is nonsense.
R9 290X memory subsystem was using 80W of power, with 6000 MHz GDDR5, with 16 memory chips.
RX 480 memory subsystem was using 37W's of power with 7000 MHz GDDR5, with 8 memory chips.

Power scales with memory bandwidth, because of the frequency. The higher frequency you go, the more power it will use.
 

raghu78

Diamond Member
Aug 23, 2012
4,093
1,475
136
Lol, power scales with bandwidth not with size...and GDDR6X has higher bandwidth efficiency.
Anyway your calculation is nonsense.

Memory power scales with bandwidth and capacity. Do you think double the VRAM on RTX Titan Ampere is going to come free of cost. Hint: The 17 Gbps on Titan Ampere vs the 21 Gbps on the RTX 3090 is for power reasons.


2nd Gen NVIDIA TITAN GA102-400-A1 5376 24GB 17Gbps
GeForce RTX 3090 GA102-300-A1 5248 12GB 21Gbps
GeForce RTX 3080 GA102-200-Kx-A1 4352 10GB 19Gbps

Here is a review of RX 5500XT 4GB and 8GB from Sapphire which are identical other than VRAM capacity. The 8GB version draws more power vs the 4GB version

 
  • Like
Reactions: Mopetar and uzzi38

Thala

Golden Member
Nov 12, 2014
1,355
653
136
Memory power scales with bandwidth and capacity. Do you think double the VRAM on RTX Titan Ampere is going to come free of cost. Hint: The 17 Gbps on Titan Ampere vs the 21 Gbps on the RTX 3090 is for power reasons.

The power adder based on capacity is mostly the power needed for refreshing the cells, which is much smaller than dynamic power required for accessing a page. So Glo's calculation continues to be nonsense.

I also have a hint for you: 8 GByte version drawing more power than the 4 GByte version is also attributed to the fact that the 8 GByte version is doing more work per time unit (e.g. games running faster).

And now go back and read again what Glo has claimed/calculated before nitpicking that power-bandwidth scaling is not totally correct.
 
Last edited:
  • Like
Reactions: Lodix

raghu78

Diamond Member
Aug 23, 2012
4,093
1,475
136
The power adder based on capacity is mostly the power needed for refreshing the cells, which is much smaller than dynamic power required for accessing a page. So Glo's calculation continues to be nonsense.

I wanted to point out that your statement about power scaling only with bandwidth and not size was incorrect. BTW you do not have a detailed breakup of memory power scaling with capacity. So one could argue you are speculating .
 

Thala

Golden Member
Nov 12, 2014
1,355
653
136
I wanted to point out that your statement about power scaling only with bandwidth and not size was incorrect. BTW you do not have a detailed breakup of memory power scaling with capacity. So one could argue you are speculating .

Nope i am not speculating, i am speaking of experience with our designs using different amount of memories, where i have precise numbers. You can also roughly calculate this by looking into the datasheet of the memories - no speculation needed.
Thing is the correlation with used bandwidth is much higher than the correlation with capacity*. Glo just multiplied power with capacity such that double the capacity took double the amount of power.

*I could imagine a combination, where you have very low effective bandwidth usage while having huge capacity, where the capacity based power starts to dominate.
 
Last edited:
  • Like
Reactions: DXDiag

IntelUser2000

Elite Member
Oct 14, 2003
8,686
3,786
136
R9 290X memory subsystem was using 80W of power, with 6000 MHz GDDR5, with 16 memory chips.
RX 480 memory subsystem was using 37W's of power with 7000 MHz GDDR5, with 8 memory chips.

Well in this case the R9 290X has double the width. It's not quite double the bandwidth, but its very close. The RX 480's memory might have an advantage of using better process as well.

Yes memory is different because capacity doesn't have a linear relation with power consumption. There's some extra power if you double the amount of chips but nowhere near double. Of course if you double capacity with same chip count the increase will be minimal.
 

Konan

Senior member
Jul 28, 2017
360
291
106
Wonder how thats going to work. Would be a strange bus layout compared to the other cards.

Sounds like the founders edition will have the irregular PCB and funky doubled sided on the card fans. The partner cards probably three slots thick and some interesting designs... maybe.. hehe
Probably some water cooled versions.

As for bus layout what do you think?
 

GoodRevrnd

Diamond Member
Dec 27, 2001
6,801
581
126
It's hard to believe because it would make these cards too good for deeplearning stuff unless NV starts blocking that on a driver level. There would be very little need to get a Titan or Quadro card which usually have more RAM.
Maybe they'll do some distributed deep learning to improve DLSS? Also, if so, spooky.
 

A///

Diamond Member
Feb 24, 2017
4,351
3,158
136
It's hard to believe because it would make these cards too good for deeplearning stuff unless NV starts blocking that on a driver level. There would be very little need to get a Titan or Quadro card which usually have more RAM.

Yeah, that does sound like NVidia. Didn't they put a stop to using GeForce in datacenter or was that words only and not actually anything that prevents the cards from being used?
 

TempAcc99

Member
Aug 30, 2017
60
13
51
Yeah, that does sound like NVidia. Didn't they put a stop to using GeForce in datacenter or was that words only and not actually anything that prevents the cards from being used?

That's words only but still legally binding. But I wasn't thinking about datacenter use but workstation use for which you can use GeForce for anything you want. The limitation with GeForce usually was the amount of RAM. If your data doesn't fit into the GPUs memory, then it gets very slow to train so that a slower quadro with more vram actually gives better performance. Hence my comment that this amount of vram is doubtful.
 

A///

Diamond Member
Feb 24, 2017
4,351
3,158
136
That's words only but still legally binding. But I wasn't thinking about datacenter use but workstation use for which you can use GeForce for anything you want. The limitation with GeForce usually was the amount of RAM. If your data doesn't fit into the GPUs memory, then it gets very slow to train so that a slower quadro with more vram actually gives better performance. Hence my comment that this amount of vram is doubtful.
You wouldn't use a single card. That's the point. NVidia's software suite would probably prevent the workload from being done if it detects GeForce cards and not Quadros. There's an open source project on the internet but it doesn't get much love.

If you're talking about straight up workstation use for rendering stuff, then certain software make use of what Quadro offers at the hardware level than what GeForce does. Quadros use ECC VRAM. If you're doing heavy renders for a mechanical device or system in software like SolidWorks a small error may really cost a lot. It's a no brainer to use Quadro at that point, including the ability to render larger files as you pointed out, simulation work and raw performance at the upper end of Quadros.

There's always going to be a major benefit to choose Quadro cards, especially if no business is going to risk messing up because they saved a few thousand dollars.