NVIDIA Maxwell Speculation Thread

[Kaotik]

You are forgetting the die space *REQUIRED* for the additional FP64 units.

If there will be such, there was a japanese article a while back in which NVIDIA rep (unless something got lost in translation) said quite clearly there won't be Maxwell Tesla-parts because of the lack of FP64 units, and that Pascal would be next Tesla, which would suggest GM200 won't include additional FP64 units either

 

[AnarchX]

Some SiSoft data:
Boosting over 1.6GHz, maybe for this are the >225W PCBs supposed:
GTX 970 with only 1.8MiB L2? Missreading or is GM2xx able to deactivate ROPs/L2 independent from IMC?

So it probably turns out, this is an additional deactivation on GTX 970s GM204, which lowers the speed for memory usage beyond 3.5GiB significantly and the driver tries to keep memory usage below 3.5GiB on GTX 970, while GTX 980 is able to fill ~4GiB.

Also the 12 SMM - 2MiB - GTX 980M is also able to fill ~4GiB: https://forums.geforce.com/default/...tx-970-3-5gb-vram-issue/post/4430975/#4430975

 

[lanek]

Is 120W really the "TDP"? Not that it really matters. Have a look at the "torture" graph: http://www.tomshardware.de/nvidia-geforce-gtx-960-maxwell-test,testberichte-241727-17.html


Granted these are done with a power virus, but I fail to see any "sweet spot" in terms of power consumption. At default clocks it stays within the 120W range. But once you let "overclocking dream" come into play it reaches nearly GTX980 level in that graph.

There's no reference gpu's who have been tested.... Nobody ( i mean no one reviewers ) have got a reference gpu... after some complaint by some site, Nvidia have send them what should have been a reference GTX960 bios for flash it on the overclocked / custom gpu's they have recieve and use that as the GTX960 reference baseline gpu for their article ...

 

[Kaotik]

Is 120W really the "TDP"? Not that it really matters. Have a look at the "torture" graph: http://www.tomshardware.de/nvidia-geforce-gtx-960-maxwell-test,testberichte-241727-17.html
Granted these are done with a power virus, but I fail to see any "sweet spot" in terms of power consumption. At default clocks it stays within the 120W range. But once you let "overclocking dream" come into play it reaches nearly GTX980 level in that graph.

You can't take into account the "custom models", reference GTX 980 TDP is 180W, but for example Gigabyte's model has it cranked up all the way up to 300W, and it shows in tests. Reference TDP is no doubt that 120W (or most likely is, have to double check reference BIOS, since 980 was advertised as 165W while it was really 180W)

 

[swaaye]

So it probably turns out, this is an additional deactivation on GTX 970s GM204, which lowers the speed for memory usage beyond 3.5GiB significantly and the driver tries to keep memory usage below 3.5GiB on GTX 970, while GTX 980 is able to fill ~4GiB.

Also the 12 SMM - 2MiB - GTX 980M is also able to fill ~4GiB: https://forums.geforce.com/default/...tx-970-3-5gb-vram-issue/post/4430975/#4430975

I sense there be some intarwebz rage brewing!!!

 

[Ailuros]

You can't take into account the "custom models", reference GTX 980 TDP is 180W, but for example Gigabyte's model has it cranked up all the way up to 300W, and it shows in tests. Reference TDP is no doubt that 120W (or most likely is, have to double check reference BIOS, since 980 was advertised as 165W while it was really 180W)

Actually I shouldn't take any reference models into account from the 960 all the way up to the 980 since in real time it's questionable if you even find one sold. In the end whether you take reference board consumption values or custom SKUs vs. custom SKUs I personally fail to see any "sweet spot" in terms of power consumption. I'd give that oscar to the GTX970 and yes that's obviously just me as I said several posts above.

 

[lanek]

I sense there be some intarwebz rage brewing!!!

http://www.guru3d.com/news-story/does-the-geforce-gtx-970-have-a-memory-allocation-bug.html

 

[Jawed]

I sense there be some intarwebz rage brewing!!!

This could be better than FX

 

[AnarchX]

So GTX 970 is actually a "3.5+0.5GB" graphics card:
http://techreport.com/news/27721/nvidia-admits-explains-geforce-gtx-970-memory-allocation-issue
http://www.pcper.com/news/Graphics-Cards/NVIDIA-Responds-GTX-970-35GB-Memory-Issue

But it why can 12 SMM - 2MiB@SiSoft - 256-Bit GTX 980M can adress the memory with full 4/8GB? Are the 13 SMM the problem? Mabye NV should avoid "such" numbers, like some asian companies. ;D

 

[ArcticCircle]

I wonder what this means for possible 8GB cards? Delayed? Not coming at all in this gen?

I think this 3.5+0.5GB situation is really sucky for developers and publishers who recently pushed graphics to new levels. Imagine if many those hitching and slowdown issues at max settings (like on Unity or Shadow of Mordor) are just because customers didn't know about this and tried to push their gear to near 4GB limit. As expected, their average fps would look somewhat ok, but they'd see the difference in frametime. It's like SLI / Crossfire frame pacing thing all over again.

I love massive textures above all else so I hope this issue gets resolved and studied properly.

 

[Jawed]

Looks like the commenters on PCPer are on the case. Whereas Scott just walks off calling it academic.

Since when have reviewers taken numbers from IHVs at face value?

And what about frame rate minima? And what about when the cause of the >3.5GB memory usage is not due to render target allocation (including MSAA, SSAA) but ultra-high quality texture packs?

 

[swaaye]

Looks like the commenters on PCPer are on the case. Whereas Scott just walks off calling it academic.

Since when have reviewers taken numbers from IHVs at face value?

And what about frame rate minima? And what about when the cause of the >3.5GB memory usage is not due to render target allocation (including MSAA, SSAA) but ultra-high quality texture packs?

Scott said look at the reviews that have already tested these cards. 970's performance is pretty well known at this point. They've been pounded with all the big titles and I don't recall anyone noticing strange behavior....

I wonder if any older products with disabled parts have similar memory performance behavior characteristics. Of course there's also the GTX 450/550 but that's not quite the same thing.

 

[MDolenc]

I would really really wish that people would stop forgetting that Windows actually use some gpu memory. Here. Win 8.1 at 1920x1080 uses 258MB gpu memory. It's not possible for "Nai's Benchmark" 3840MB on a 4GB board and make it all fit into vram. Depending on resolution, number of displays, amount of crap running (browsers) the amount of memory can shrink further or it can be lower if you use Win 7 for example. And some cuda buffers that this nice banchmark is using are overflowing to system memory.

 

[Paran]

I would really really wish that people would stop forgetting that Windows actually use some gpu memory. Here. Win 8.1 at 1920x1080 uses 258MB gpu memory. It's not possible for "Nai's Benchmark" 3840MB on a 4GB board and make it all fit into vram. Depending on resolution, number of displays, amount of crap running (browsers) the amount of memory can shrink further or it can be lower if you use Win 7 for example. And some cuda buffers that this nice banchmark is using are overflowing to system memory.

It is possible. Boot up with your iGPU and run the benchmark. This is what clever people did. There is exactly 0MB in use on my GTX 970 when I do this.

 

[MDolenc]

Exactly. And what do you get then from this benchmark?

 

[Paran]

http://fs1.directupload.net/images/150124/h9l5bvxd.png

I get this.

 

[mczak]

I guess this should give some good hints about how SMMs and MCs are actually connected? I can't make much sense out of it though, the numbers sort of seem to imply the mapping between MCs and SMMs are completely fixed so performance for the memory which can't be mapped that way drops to rock bottom. But this can't really be...

 

[Kaotik]

I guess this should give some good hints about how SMMs and MCs are actually connected? I can't make much sense out of it though, the numbers sort of seem to imply the mapping between MCs and SMMs are completely fixed so performance for the memory which can't be mapped that way drops to rock bottom. But this can't really be...

The way I understood it is that the problem is in the crossbars handling communication between SMMs and MCs, not in MCs nor SMMs

 

[Grall]

All of this could easily have been avoided by calling the 970 a 3.5GB product even though it has (and can use) 4GB onboard and being forthright about the performance issue behind this decision, but nooo, as usual, NV had to be greedy and shady. Truly, they're one of the most amoral entities in the PC hardware business.

They didn't think someone would figure this out on their own? Well, why keep it a secret? Because their product looks better if they STFU about it, of course. Essentially, this is a bait-and-switch, of a fashion.

Not the first time they pull a stunt in this fashion. Is it in Jen-Hsun's blood to cheat people whenever he thinks he can get away with it?

 

[AnarchX]

I wonder what this means for possible 8GB cards? Delayed? Not coming at all in this gen?

If they use same die configuration like on current GTX 970s, you get a 7GB + 1GB card.

I guess this should give some good hints about how SMMs and MCs are actually connected? I can't make much sense out of it though, the numbers sort of seem to imply the mapping between MCs and SMMs are completely fixed so performance for the memory which can't be mapped that way drops to rock bottom. But this can't really be...

On 12 SMM - 2MiB@SiSoft - 256-Bit GTX 980M you get a full 4/8GB memory pool. Its probably this additional deactivation of L2-Cache on GTX 970 GM204-200, which brings this problem. SiSoft read through CUDA from the first day, that there are only 1.8MB L2 on GTX 970.