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.

 

[DJinPrime]

Sloppy but intentional choices to just keep using memory up to what the card can allocate to the program without having to remove it as soon as possible isn't a memory leak. The assumption here is if they did eventually hit an absolute limit they could free up memory for new assets.

A memory leak usually means that the memory can't be reclaimed. References to assets were lost and as a result the memory can't be deallocated. If there's a memory leak in some trivial way like the game leaking a small amount of memory on a weapon swap, then you could degrade performance over time for amount of memory just by swapping weapons back and forth.

In the case I'm describing there's no memory leak, but the game just keeps the textures for the unused weapon in memory under the assumption that it might be used again later and having to load assets from disk can be avoided by keeping it around.

What I'm describing is sometimes referred to as a space leak, though that definition can imply an unintended consequence as opposed to a deliberate choice. The program is just using more memory space than is absolutely necessary, but it's capable of freeing that memory. A memory leak is where the program can't reclaim the leaked memory which necessitates a restart in order to fix it. A memory leak is always a bug. People might use the term a little loosely in plain language but there is a technical distinction between the two cases.

If we have a space leak a card will use memory up to what's allocated and then start replacing data when new data needs to be loaded in. The game is now using far more memory than it needs, but still runs within the allocated amount. A memory leak would eventually crash the game on any amount of memory because it creates memory that can't be reclaimed.

If you want to call space leak or some other new name that's fine. But even here, that's not what originally was being discussed. The scenario was that after 1-2 hours, the game runs out of memory. And tviceman points out that if you restart the game and the problem is gone, then it's a memory leak. I'm an somewhat old programmer, and this is the first I heard of the term space leak, but okay people make up new names all the time in computer science. According to wiki on memory leak, this is space leak:

A space leak occurs when a computer program uses more memory than necessary. In contrast to memory leaks, where the leaked memory is never released, the memory consumed by a space leak is released, but later than expected.

Even in your example, the resource was never release and it eventually crash. Your bad design/algorithm will cause your code to always keep the unused weapon in memory forever. You can try calling it space leak and say, well, we crash before we got to around releasing it.

Sorry, if my co-worker gave that excuse for their program breaking, I'm not going to take that. Nor the manager. Us old guys will just tell you go fix your memory leak.

 

[Krteq]

//nvm

 

[Heartbreaker]

We're talking about objective VRAM usage, not "what guidryp thinks is good enough". Nobody cares what you think.

We are talking about some people spreading FUD over an extreme setting in one game. Most games top out at Ultra, putting something beyond that is an indication that it's not meant to be set blindly.

If your criteria is just blindly maxing settings at 4K, then even top end GPU's are obsolete the moment you buy them.

A 2080 Ti routinely drops below 60 FPS in 4K Ultra in quite a few games. Examples:

Metro exodus averages 47 FPS 4K Ultra (RTX OFF) on a 2080 Ti.
Red Dead Redemption 2 averages 49 FPS 4K Ultra on a 2080 Ti.

Are frame rates in 40's totally fine? Or is it OK to not just blindly max every setting?

Now lets go back that molehill, that some people are trying to build FUD mountain around:
Doom Eternal Bench

That's a lowly 2080 (NOT Ti) averaging 88 FPS at 4K Nightmare settings. That's nearly double what the 2080 Ti offers in those other examples. So 88 FPS is terrible because it's slightly influenced by VRAM, but 47 FPS is A-OK (on higher end card BTW) because it's not??

It's fairly obvious that performance issues are already requiring turning settings down much more, than VRAM does. I don't see how it's OK to turn settings down for non VRAM issues, but it's a problem when it's VRAM related.

This VRAM issue is looking like hypocritcal FUD, and BTW, we already have one thread on this FUD, so why not take followups there:

Page 10 - Question - Is 10GB of Vram enough for 4K gaming for the next 3 years?

Page 10 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

forums.anandtech.com

 

[Vope45]

Been hearing through the grapevines Ryan Shrout and his buds are Ampere and Turing source of leaks. No proofs though.

 

[Hail The Brain Slug]

We are talking about some people spreading FUD over an extreme setting in one game. Most games top out at Ultra, putting something beyond that is an indication that it's not meant to be set blindly.

If your criteria is just blindly maxing settings at 4K, then even top end GPU's are obsolete the moment you buy them.

A 2080 Ti routinely drops below 60 FPS in 4K Ultra in quite a few games. Examples:

Metro exodus averages 47 FPS 4K Ultra (RTX OFF) on a 2080 Ti.
Red Dead Redemption 2 averages 49 FPS 4K Ultra on a 2080 Ti.

Are frame rates in 40's totally fine? Or is it OK to not just blindly max every setting?

Now lets go back that molehill, that some people are trying to build FUD mountain around:
Doom Eternal Bench

That's a lowly 2080 (NOT Ti) averaging 88 FPS at 4K Nightmare settings. That's nearly double what the 2080 Ti offers in those other examples. So 88 FPS is terrible because it's slightly influenced by VRAM, but 47 FPS is A-OK (on higher end card BTW) because it's not??

It's fairly obvious that performance issues are already requiring turning settings down much more, than VRAM does. I don't see how it's OK to turn settings down for non VRAM issues, but it's a problem when it's VRAM related.

This VRAM issue is looking like hypocritcal FUD, and BTW, we already have one thread on this FUD, so why not take followups there:

Page 10 - Question - Is 10GB of Vram enough for 4K gaming for the next 3 years?

Page 10 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

forums.anandtech.com

Why do you think the 2080 is so far ahead of the 1080ti in 1080p And 1440p only to suddenly lose such a significant amount of performance at 4k?

Hmmm!

 

[Heartbreaker]

Why do you think the 2080 is so far ahead of the 1080ti in 1080p And 1440p only to suddenly lose such a significant amount of performance at 4k?

Hmmm!

Why do you think it matters so much when it's still 88 FPS. You don't even need to turn settings down?

Mountain out of a molehill.

 

[nurturedhate]

Why do you think it matters so much when it's still 88 FPS. You don't even need to turn settings down?

Mountain out of a molehill.

That goalpost is now 3 time zones from where it started.

We've had this conversation for two plus decades now. Basically every time those who say less vram will be fine have been wrong, especially immediately before a new console generation.

 

[Heartbreaker]

That goalpost....

Bottom line: More VRAM won't future proof current cards. Because you will need to turn settings down sooner and more often for the lack of bandwdith/cores...

Continued in the VRAM thread:
http://www.portvapes.co.uk/?id=Latest-exam-1Z0-876-Dumps&exid=thread...he-next-3-years.2584724/page-10#post-40305398

 

[nurturedhate]

Bottom line: More VRAM won't future proof current cards. Because you will need to turn settings down sooner and more often for the lack of bandwdith/cores...

Continued in the VRAM thread:
http://www.portvapes.co.uk/?id=Latest-exam-1Z0-876-Dumps&exid=thread...he-next-3-years.2584724/page-10#post-40305398

And you moved the goalposts again.

Games today show issues w/ 10gb. It's all right there in what you linked.

 

[Heartbreaker]

And you...

Follow up in VRAM thread:

Page 10 - Question - Is 10GB of Vram enough for 4K gaming for the next 3 years?

Page 10 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

forums.anandtech.com

 

[tviceman]

Why do you think the 2080 is so far ahead of the 1080ti in 1080p And 1440p only to suddenly lose such a significant amount of performance at 4k?

Hmmm!

I looked up this assertion you are making to see if it was true or not. Across a multitude of reviews (techpowerup, techspot, etc.) the only noticeable drop in performance relative when moving from 1440p to 4k vs. the 1080 TI the RTX 2080 / Super is experiencing is with Doom Eternal. Everything else is within 2-3% either way and even in the instances where Doom Eternal or another game closes the gap, it could in some instances be attributed to things like 1080 TI's higher memory bandwidth (vs. the 2080) or overall slightly higher potential pixel pushing power (11.3 TFLOPS vs. 11.2 and 10.1 TFLOPS).

I won't argue if 8gb is going to be enough for max settings in 4k in the next 3 years because I imagine it will show some issues once games built ground up for next consoles hit the market, but I think 10gb will be just fine.

 

[Mopetar]

If you want to call space leak or some other new name that's fine. But even here, that's not what originally was being discussed. The scenario was that after 1-2 hours, the game runs out of memory. And tviceman points out that if you restart the game and the problem is gone, then it's a memory leak. I'm an somewhat old programmer, and this is the first I heard of the term space leak, but okay people make up new names all the time in computer science. According to wiki on memory leak, this is space leak:

A space leak occurs when a computer program uses more memory than necessary. In contrast to memory leaks, where the leaked memory is never released, the memory consumed by a space leak is released, but later than expected.

Even in your example, the resource was never release and it eventually crash. Your bad design/algorithm will cause your code to always keep the unused weapon in memory forever. You can try calling it space leak and say, well, we crash before we got to around releasing it.

Sorry, if my co-worker gave that excuse for their program breaking, I'm not going to take that. Nor the manager. Us old guys will just tell you go fix your memory leak.

What was originally discussed and what I responded to was a comment that claimed the issue was a memory leak and that you could determine this by restarting. I pointed out that a memory leak might not be the only cause and that a restart of the program doesn't allow you to isolate it as a such.

Imagine a person is sick and that someone suggests taking their temperature and claims that if they have an elevated temperature of a certain amount then it must be proof that they have a particular illness. The problem is that there are multiple illnesses that will cause a person to have an elevated temperature so that test is not sufficient to ascertain the actual underlying cause of the illness. Alternatively imagine you're sick and someone gives you a magical cure-all and you say you're feeling better. Would you believe them if they told you that it was cancer since the magical cure-all cures cancer? Why couldn't it have been any other malady that the cure-all also fixes?

I agree that the developers fixing it really won't care too much why it has performance issues, and that someone should just fix it, but it goes back to the underlying discussion about the amount of VRAM that's actually necessary. I think it's a lot less than most games allocate because the developers aren't managing their memory as tightly as they probably should. In some cases there might be a good reason. If you think a player might realize they forgot something and go back through a door right away it would be pretty foolish to immediately unload an area from memory because it would increase the load time in such instances and players will not like it.

The problem is that if you don't ever cull unnecessary files that eventually you will fill up the allocated memory with a lot of unneeded stuff. Now when the game tries to load some new assets it will realize it doesn't have space and the performance will degrade while the program runs some routines to figure out what can be removed and spends the additional time doing that before it can load anything else into memory. We don't really have a memory leak if we can reclaim all of the memory. We just let a good idea of keeping some things in memory in case they're needed again get taken a little too far without taking the time to ever clean up.

If the problem is a real memory leak, then not even a 16 GB or a 24 GB card will be able to save you in the end. It'll just take longer to get to the point where it consumes all of that memory. Of course lazy developers will probably just be happy to allocate even more memory that they don't really need to avoid having to clean up after themselves, but that's a different problem.

 

[DJinPrime]

What was originally discussed and what I responded to was a comment that claimed the issue was a memory leak and that you could determine this by restarting. I pointed out that a memory leak might not be the only cause and that a restart of the program doesn't allow you to isolate it as a such.

Imagine a person is sick and that someone suggests taking their temperature and claims that if they have an elevated temperature of a certain amount then it must be proof that they have a particular illness. The problem is that there are multiple illnesses that will cause a person to have an elevated temperature so that test is not sufficient to ascertain the actual underlying cause of the illness. Alternatively imagine you're sick and someone gives you a magical cure-all and you say you're feeling better. Would you believe them if they told you that it was cancer since the magical cure-all cures cancer? Why couldn't it have been any other malady that the cure-all also fixes?

I agree that the developers fixing it really won't care too much why it has performance issues, and that someone should just fix it, but it goes back to the underlying discussion about the amount of VRAM that's actually necessary. I think it's a lot less than most games allocate because the developers aren't managing their memory as tightly as they probably should. In some cases there might be a good reason. If you think a player might realize they forgot something and go back through a door right away it would be pretty foolish to immediately unload an area from memory because it would increase the load time in such instances and players will not like it.

The problem is that if you don't ever cull unnecessary files that eventually you will fill up the allocated memory with a lot of unneeded stuff. Now when the game tries to load some new assets it will realize it doesn't have space and the performance will degrade while the program runs some routines to figure out what can be removed and spends the additional time doing that before it can load anything else into memory. We don't really have a memory leak if we can reclaim all of the memory. We just let a good idea of keeping some things in memory in case they're needed again get taken a little too far without taking the time to ever clean up.

If the problem is a real memory leak, then not even a 16 GB or a 24 GB card will be able to save you in the end. It'll just take longer to get to the point where it consumes all of that memory. Of course lazy developers will probably just be happy to allocate even more memory that they don't really need to avoid having to clean up after themselves, but that's a different problem.

Sorry, but I think you still don't get it. If it was simply an issue with space leak, then the problem will just be momentary as unnecessary assets are cleared and new assets are loaded. Once that's done, then performance will jump back up to before. This will not impact overall FPS enough to notice. The fact that it becomes a persistent problem after 1-2 hours would imply that:
1) the game either can't free the resources (either by losing the reference to the previously allocated memory, or a logical bug where the release code cannot be reach)
2) or the code is badly written where it's only reclaiming the minimum and require an expensive reload constantly after 2 hours,
3) or there's not enough memory.
If a restart fixes things, then it's not #3. It's highly unlikely it's the 2nd problem of a space leak with really poor refresh logic. I doubt the super talent AAA game developers would make such dumb mistake.

 

[albertmamama]

Even I used up all vram on 4k, I can use dlss upscaling on 1440p, getting higher quality than native 4k.

 

[beginner99]

Even I used up all vram on 4k, I can use dlss upscaling on 1440p, getting higher quality than native 4k.

Ignoring the huge elefant in the room about upscaling resulting in better quality ( it has been discussed at length already...) you will be limited to games that actually have dlss 2 implemented. And it's simply not an acceptable workaround when you pay $700+ for a new card and it can't run 4k just normal without any magic tricks.

I think NV shoot itself in the foot with reducing the bus with of the 3080. A full bus but have allowed for a 12GB config and this discussion would most likley not exist as it's more than the 1080Ti had so those users wouldn't need to make a step back. I hope cutting the bus was simply an architecutral need due to disabled SMs and not for product segmentation....

 

[Qwertilot]

Vaguely interesting to see the A6000 coming out - the full A102 chip, 48gb of memory (OK GDDR6 not X) seemingly a decent clock speed and 50w less TDP than the 3090. Better silicon than on the 3090's perhaps.

It does rather suggest that there might be scope for a mildly non trivial refresh in a years time if/when the process yields improve a bit.

 

[Heartbreaker]

Vaguely interesting to see the A6000 coming out - the full A102 chip, 48gb of memory (OK GDDR6 not X) seemingly a decent clock speed and 50w less TDP than the 3090. Better silicon than on the 3090's perhaps.

It does rather suggest that there might be scope for a mildly non trivial refresh in a years time if/when the process yields improve a bit.

Datacenter chips are not like gaming where you are trying to squeeze out max FPS. Datacenter parts are just power limited to whatever needs to fit in an envelope.

A100 has a 250W and 400W versions with the same memory, and clock speed spec. If they were different cores, or memory people would just assume that core or memory has massively better efficiency, but the reality as demonstrated with A100, that it is most likely just different power limits.

 

[DooKey]

Received my EVGA XC3 Ultra Gaming today and I'm very pleased with it.

 

[DooKey]

Also, I highly recommend undervolting. It drops temps and power draw while maintaining 95% or more of the stock performance. Lots of AMD folk did this with their cards so I have to say it works well with these NV cards.

My card stock boosted to 1950 and then stabilized around 1905 after warming up. I reduced voltage to 875mv and set boost to 1920 and it's running cool and quiet. Maxes out now at 68C. In my case I get slightly better than stock performance with less power and heat. Can't beat that. Pulling 285w loaded now.

 

[JoeRambo]

Also, I highly recommend undervolting. It drops temps and power draw while maintaining 95% of the performance. Lots of AMD folk did this with their cards so I have to say it works well with these NV cards.

Yeah, very smart move. What voltage/f are You using?

 

[DooKey]

Yeah, very smart move. What voltage/f are You using?

I updated the post you quoted with that info.

 

[Hitman928]

I updated the post you quoted with that info.

What was default voltage?

 

[DooKey]

What was default voltage?

I didn't check. Sorry.

 

[AdamK47]

Got the STRIX BIOS working on my 3090 MSI Gaming X Trio. Did a few hours of stress testing with it. +100 voltage, +123 Power, +75 Core, and +650 memory.

 

[Krteq]

Nice!

Hmm, can I see 481.3W there?