Idea: should nvidia offer a way to run the 970 as just 3.5 GB?

[DaveSimmons]

I posted this in the compensation thread but it's a separate topic.

People point out that the 970 works great for today's games, and that you have to work to trigger memory use above 3.5 GB that causes problems.

What if nvidia offers a driver or firmware update that lets you disable the slow 512 MB?

With that turned on, it becomes a 3.5 GB card.

Con: when the driver/firmware code works, the extra 512 MB might work well as a RAMdisk type high-speed cache for little used resources.

Pro:

it takes away uncertainty. The 970 becomes a normal 3.5 GB card. No driver or firmware tricks are needed to prevent memory access stalls.

There are no worries about future games accidentally using the "bad" 512 MB before they need to because nvidia stops writing game-specific fixups for the 970 into the drivers.

 

[tential]

So when I need 4GB of VRAM and I only have 3.5 GB of VRAM, it would rather than use all my VRAM at 4GB, use only 3.5 GB of VRAM and system ram?

Isn't that slower?

 

[ocre]

I said this on multiple threads

Nvidia mod on their site claims their driver team is gonna revamp the memory management. Said there are many things other than game textures in VRAM, such as OS, apps, and driver. He said they are working on this.

I know that stuff can take 300mb with just windows and nothing open. It can go up to 500mb, as many people might know. So it actually a promising driver change. See, the 980 has stuff loaded in its ram that is not game related at all times. If nvidia can successfully stuff this segment up and keep the 3.5 gb for games, the penalty can be almost completely reduced in gaming. How practical this is, idk. Just heard they are working on this driver.

We just have to see how it works

So, OP

Nvidia is supposedly gonna make some changes that sound promising

 

[xthetenth]

So when I need 4GB of VRAM and I only have 3.5 GB of VRAM, it would rather than use all my VRAM at 4GB, use only 3.5 GB of VRAM and system ram?

Isn't that slower?

I think the idea is a flag to set so a game only sees 3.5 GB and doesn't do stuff like load up higher LoD assets than needed thinking it's got 4GB of memory to use and mixing in needed things into the upper .5 GB, dropping bandwidth considerably in the proccess as it starts not being able to read whatever it needs from the memory at the same time.

I think it'd be a good way to get rid of my worst fear about holding onto this card for a while which is driver support drying up and leaving me with a worse than 3.5 GB card in modern games in a few years.

 

[wand3r3r]

@OP Yes, imo.

So when I need 4GB of VRAM and I only have 3.5 GB of VRAM, it would rather than use all my VRAM at 4GB, use only 3.5 GB of VRAM and system ram?

Isn't that slower?

You would probably be forced to cut back a setting and hit a normal hard limit.

The way that they have done it seems strange to me, the whole VRAM is slowed down when using the last 512 MB and therefore the whole issue.

I would prefer 3.5 GB for gaming as a hard limit. If they could get the extra slow VRAM to be used for windows or something else that would just be a bonus cache.

 

[Eymar]

Depends. If future cards have this memory configuration then I wouldn't really be worried about driver issues for next 2-3 years as the fixes\performance improvements should be common. If the 970 is unique and driver fixes have longer turn around time, then I'd think 3.5GB limit is preferable.

 

[flexy]

I said this on multiple threads

Nvidia mod on their site claims their driver team is gonna revamp the memory management. Said there are many things other than game textures in VRAM, such as OS, apps, and driver. He said they are working on this.

I know that stuff can take 300mb with just windows and nothing open. It can go up to 500mb, as many people might know. So it actually a promising driver change. See, the 980 has stuff loaded in its ram that is not game related at all times. If nvidia can successfully stuff this segment up and keep the 3.5 gb for games, the penalty can be almost completely reduced in gaming. How practical this is, idk. Just heard they are working on this driver.

We just have to see how it works

So, OP

Nvidia is supposedly gonna make some changes that sound promising

This sounds indeed interesting although I ASSUMED they're already doing something like that in the drivers/BIOS.

However, *IF* there is headroom to shove allocated memory around and "park it" (so to speak) in the slower memory location this could ALMOST solve all the problems for the 970. Fact is that indeed a good chunk of video memory is allocated by Windows...so this would otherwise "waste" the good/fast memory which of course should be available to games only.

 

[SimianR]

From my understanding the drivers already did this to some extent. That's how the problem was initially discovered, people were finding that the vram usage was capping out at 3500mb unless they pushed the settings further and further.

 

[Spanners]

I think assuming the 0.5GB is managed well it should be beneficial, the problems arise when it's not managed well and then performance may well be worse than it not being accessible at all, from a smoothness point of view anyway.

From a business/liability standpoint I'd be extremely surprised if they went down this road.

The positive thing for owners in all this is the spotlight should compel Nvidia to put more resources towards improving the way drivers handle the 0.5GB. Which it seems they are already doing.

 

[Magic Carpet]

Nvidia has been compromised.

Drivers won't help when nvidia stops caring about maxwell. It's best to get a better design hardware, imo. At any case, 970 will have a poorer resale value. Why bother.

This reminds me of an iphone issue, when Jobs said, the phone is fine is just you need to carry it in a specific way. Eventually, issuing out free cases for the affected users.

Instead of admitting a design compromise, they are saying, there is no problem, our drivers team will take care of everything, worry not. Yeah, you tell that the Titan owners who spent thousands of dollars and now are being neglected in the driver department. Way to go, Nvidia :thumbsdown:

I hope, AMD remains in business.

 

[BFG10K]

What if nvidia offers a driver or firmware update that lets you disable the slow 512 MB?

What for? Accessing that 512MB segment is still faster (with less latency) than PCIe or system memory.

By disabling it, performance will drop even further overall.

 

[JoeRambo]

Nvidia mod on their site claims their driver team is gonna revamp the memory management. Said there are many things other than game textures in VRAM, such as OS, apps, and driver. He said they are working on this.

I know that stuff can take 300mb with just windows and nothing open. It can go up to 500mb, as many people might know. So it actually a promising driver change. See, the 980 has stuff loaded in its ram that is not game related at all times. If nvidia can successfully stuff this segment up and keep the 3.5 gb for games, the penalty can be almost completely reduced in gaming. How practical this is, idk. Just heard they are working on this driver.

This is actually very dangerous. A lot of people now run windowed mode games and multi monitors, such change like showing windows stuff into 512MB will have disastrous performance for the game even if it is not using full 3.5GB. GTX970 is 24-196GB/s memory bw card, where every access in last 512MB block is hurting performance.

What for? Accessing that 512MB segment is still faster (with less latency) than PCIe or system memory.

By disabling it, performance will drop even further overall.

Any access is lower latency, but kills bandwith on the other segment, causing frame time spikes, at some point it is wiser to go PCIE and then read from the main 3.5GB block. Note that to read 0.5GB you need ~15ms @28GB/s, while the rest of card stalls.

 

[DaveSimmons]

If this was a toggle option, you could leave it off until there is some game causing problems, or some Windows update that breaks the driver's attempts to manage memory to prevent lag spikes.

 

[exar333]

I would like to see a detailed review comparing 3.5GB vs. 'driver-optimized' 3.5GB+0.5GB to make the right decision.

 

[Riceninja]

that nvidia mod has backtracked and since stated that they are -not- coming out with a 970 specific update:

https://forums.geforce.com/default/topic/803518/geforce-900-series/gtx-970-3-5gb-vram-issue/219/

 

[Pneumothorax]

that nvidia mod has backtracked and since stated that they are -not- coming out with a 970 specific update:

https://forums.geforce.com/default/topic/803518/geforce-900-series/gtx-970-3-5gb-vram-issue/219/

This crap totally reeks of a mid-level admin who wanted to help the 970 owners initially, but once the 'higher-up's' and 'bean counters' got a hold of the situation in turn directed the lower level guys with this new 'policy'.

serious backtracking here Nvidia...

 

[mindbomb]

you can control the memory usage yourself with the texture quality settings in games. Just choose a setting that will never go above 3.5gb.

 

[Vesku]

Wouldn't hurt to have a driver setting to switch between 3.5GB and 4GB. Would certainly help people troubleshoot issues with a game and exclude that extra 512MB of slow memory as the particular culprit.

Edit: Plus it would benefit people using it for certain GPU Compute tasks or whose non-game programs don't like that slower section of VRAM.

 

[BFG10K]

Any access is lower latency, but kills bandwith on the other segment, causing frame time spikes, at some point it is wiser to go PCIE and then read from the main 3.5GB block. Note that to read 0.5GB you need ~15ms @28GB/s, while the rest of card stalls.

The two partitions are still accessible at once, and nVidia claims 4%-6% hit overall when both are. In order to prove the PCIe route is better, you'd need to run the 470 with 3.5 GB capped as a comparison.

If you're above 3.5GB, expect that 512MB segment to be swapped over PCIe based on the LRU policies the driver/OS/GPU use, so in a streaming game it could be constant.

I'd bet with 95% confidence the current split configuration would work better, and I wouldn't be surprised if nVidia at least simulated it and discovered the same thing.

 

[altairauditore]

Wait a minute:

Does accessing the slow 0.5GB partition slow the entire memory? If so, it would be useful to have an option to disable it. There seems to be a disparity in data provided by different sites relating to this.
If it does not, why do it?

The big question is how much of an effect the slow partition creates. Is it better to have 3.5GB VRAM and make the GPU rerender the scene or textures, or is it better to keep the stuff in the 0.5GB partition and access it from there?

I feel Nvidia should have a driver setting that allow you to disable it if you want to.

 

[skipsneeky2]

May be great to disable it,i run OSD when i game often times so i adjust settings as needed to keep from hitting the vram limit.Have had my 770 long enough to know with my current games which settings play well fps and which don't cap my vram.

Guess the 970 should be simple,keep to 1080p and simply keep to my minimums of 60+ as i have done with my 770 and i shouldn't have a problem for quite a while if at all with the 970.

 

[coercitiv]

Wait a minute:

Does accessing the slow 0.5GB partition slow the entire memory?

Read this, taken from the AnandTech's article. Once Ryan Smith said it and did not get corrected by Nvidia... my guess is this is true.

NVIDIA has segmented the GTX 970’s memory into the now-familiar 3.5GB and 512MB segments. In the case of the 3.5GB segment, this behaves otherwise identically to a fully enabled card such as the GTX 980, with the 1KB stride being striped over 7 crossbar ports, and hence 7 DRAM modules. Meanwhile the 8th and final DRAM module sits in its own 512MB segment, and must be addressed by the crossbar on its own.

This in turn is why the 224GB/sec memory bandwidth number for the GTX 970 is technically correct and yet still not entirely useful as we move past the memory controllers, as it is not possible to actually get that much bandwidth at once when doing a pure read or a pure write. In the case of pure reads for example, GTX 970 can read the 3.5GB segment at 196GB/sec (7GHz * 7 ports * 32-bits), or it can read the 512MB segment at 28GB/sec, but it cannot read from both at once; it is a true XOR situation. The same is also true for writes, as only one segment can be written to at a time.

 

[JoeRambo]

The two partitions are still accessible at once, and nVidia claims 4%-6% hit overall when both are. In order to prove the PCIe route is better, you'd need to run the 470 with 3.5 GB capped as a comparison.

The information on the web about this is conflicting at best. But somehow i doubt that NV overbuilt crossbar/L2 to handle this situation, sounds like waste of silicon, unless they were planning this exact scenario.

How can one reconcile this (source: PCPER):

If you walked across the memory interface in the exact same way, over the same 4GB capacity, the 7th crossbar port would tend to always get twice as many requests as the other port (because it has two memories attached). In the short term that could be ok due to queuing in the memory path. But in the long term if the 7th port is fully busy, and is getting twice as many requests as the other port, then the other six must be only half busy, to match with the 2:1 ratio. So the overall bandwidth would be roughly half of peak.

with this:

UPDATE 1/27/15 @ 5:36pm ET: I wanted to clarify a point on the GTX 970's ability to access both the 3.5GB and 0.5GB pools of data at the same. Despite some other outlets reporting that the GPU cannot do that, Alben confirmed to me that because the L2 has multiple request busses, the 7th L2 can indeed access both memories that are attached to it at the same time.

 

[xthetenth]

This crap totally reeks of a mid-level admin who wanted to help the 970 owners initially, but once the 'higher-up's' and 'bean counters' got a hold of the situation in turn directed the lower level guys with this new 'policy'.

serious backtracking here Nvidia...

Seriously making my future purchasing decisions easy. This went from annoying but well handled to badly handled pretty quickly.

 

[skipsneeky2]

3.5gb 970/4gb 980
2gb 770/ 3gb 780
1.2gb 570/ 1.5gb 580
896mb 275/1gb 280
512mb 8800gt/768mb 8800gtx

The above and the whole 3.5gb 970 issue suddenly really isn't surprising me much honestly,expecting a true 4gb 970 would have broken a trend Nvidia has been following way back to 8800 series if not even longer on their top tier and second best cards.

It still sucks but i wouldn't be surprised if this trend continues going forward in the future.