System Light Dark

xentr_theme_editor

xentr_page_setup xentr_appearance xentr_sidebar

  • xentr_page_width

    xentr_toggle_page_width

    xentr_page_width_description

    xentr_color_pickers

    xentr_toggle_color_picker

    xentr_toggle_color_picker_description

    xentr_typography

    xentr_select_the_font_you_want_to_use

    xentr_node_layout

    xentr_grid_layout

    xentr_grid_layout_description

  • Style variation

    xentr_select_the_style_variation_you_want_to_use
    System Light Dark

  • xentr_sidebar

    xentr_toggles_sidebar

    xentr_sidebar_description

    xentr_sticky_sidebar

    xentr_sticky_sidebar_description
    xentr_sidebar_position
    Left
    Right
  • Please do not post any links until you have 3 posts as they will automatically be rejected to prevent SPAM. Many words are also blocked due to being used in SPAM Messages. Thanks!

The GTX 970's Memory Explained & Tested (Comment Thread)

Sagath said:
So wait. I want to be sure I have this right...A card with 81.25% of the TMU's that an 980 has is getting 80% of the performance of a gtx980 at 1440p?

Well then...allow me to retort: As you might know, performance never scales perfectly (1:1) with extra resources, so even worst case we're looking at a perfect example of why this card is performing 20% slower then a card with 20% more TMUs, which is pritty damn ******* close to 1:1, and doesnt show a goddamn thing about it being the memory partition as the culprit, but probably being almost exactly 1:1 performance limited on TMU's and ROP's by 20%. :whistle:
Click to expand...

You do not understand the meaning of that statement. It means when you add resources don't expect perf to scale 1:1. So if you keep GTX 980 as 1x TMU resources then GTX 970 = 0.8125. Alternatively if you keep GTX 970 as 1x TMU then GTX 980 has 1.23x the TMU resources. So in reality even if TMU is bottlenecking the performance and there is a 1:1 perf correlation which never happens btw the real world perf cannot be lower than 0.8125x. But in this case it does. GTX 970 perf is 0.8x if GTX 980 is kept as 1x and if GTX 970 is kept as 1x then GTX 980 is 1.25x. Its easy to see that TMU alone is not the factor at play here and the memory partition is coming to play here.

YukonTrooper said:
Those charts say to me the memory bus is coming into play, rather than the memory size. The 290, 780 Ti and Titan all end up surpassing the 970 on the high resolution benchmark, and the 290X surpasses both the 970 and 980. Furthermore, if the memory size was coming into play, you'd expect all the 3GB cards to plummet, but the 780 Ti stays strong and the rest of the 3GB cards gain considerably, including the 780 coming to within a couple FPS of the 970.
Click to expand...

yeah that 196 GB/s bandwidth for the first 3.5 GB and 28 GB/s for the last 0.5 GB is at play even though the people who defend GTX 970 are saying its not.
 
Last edited:
Fragman said:
yeah that 196 GB/s bandwidth for the first 3.5 GB and 28 GB/s for the last 0.5 GB is at play even though the people who defend GTX 970 are saying its not.
Click to expand...
What I'm saying is, at least in those graphs, it's the narrow memory bus of the 970/980 that looks to be the culprit, not the memory size (or slower 0.5GB partition on the 970).
 
YukonTrooper said:
Those charts say to me the memory bus is coming into play, rather than the memory size. The 290, 780 Ti and Titan all end up surpassing the 970 on the high resolution benchmark, and the 290X surpasses both the 970 and 980. Furthermore, if the memory size was coming into play, you'd expect all the 3GB cards to plummet, but the 780 Ti stays strong and the rest of the 3GB cards gain considerably, including the 780 coming to within a couple FPS of the 970.
Click to expand...

This is exactly why it cannot be the memory bus actually. If the 780Ti stays strong with 4GB of textures loaded, then 1GB of textures are being cached off card at 22GB/s, instead of its memory bus coming in to play at 336gb/s. Whereas the 970 would have 3.5GB running at (196) and another 512 running at (28), and none at all running at 22GB/s cached off card.

The math doesnt work.

Fragman said:
You do not understand the meaning of that statement. It means when you add resources don't expect perf to scale 1:1. So if you keep GTX 980 as 1x TMU resources then GTX 970 = 0.8125. Alternatively if you keep GTX 970 as 1x TMU then GTX 980 has 1.23x the TMU resources. So in reality even if TMU is bottlenecking the performance and there is a 1:1 perf correlation which never happens btw the real world perf cannot be lower than 0.8125x. But in this case it does. GTX 970 perf is 0.8x if GTX 980 is kept as 1x and if GTX 970 is kept as 1x then GTX 980 is 1.25x. Its easy to see that TMU alone is not the factor at play here and the memory partition is coming to play here.


yeah that 196 GB/s bandwidth for the first 3.5 GB and 28 GB/s for the last 0.5 GB is at play even though the people who defend GTX 970 are saying its not.
Click to expand...

Your math doesnt add up. Well, actually it does, but your statement does not. The ratios off of a same architecture card are almost perfectly in line with a penalty expected for having less resources.

Also cards with less RAM then then 970 (if it was the RAM causing the issue) would experience a PCIE bottleneck of 22 GB/s faster then the 970, so I still dont see any where that you or the article have explained why this is the culprit for the benchmark dip.

Wouldnt this be a really easy test to prove by simply overclocking the RAM on the 780/970/980 and rerunning the tests?
 
Fragman said:
Its easy to see that TMU alone is not the factor at play here and the memory partition is coming to play here.
Click to expand...

Man you are really pushing to try and "prove" that this is a much larger issue than it is. I understand it is an issue, but every time the 970 falters it doesn't automatically mean its the freaking 0.5gb partition. The linked reviewer seems to want to push that flaw as well. Just look at the chart with open eyes for one minute....not just at the 970 and 980. If you do that you will see a 3GB 780Ti that's got better frame rates than the 970.

Why is that important?....I am glad you asked. If the game was running more than 3gb of vram as tested the 780Ti would falter to much lower performance due to insufficient vram. Because that is not the case it more than likely means that the 970 is no where near using its full "good 3.5gb" let alone having to run into its 0.5gb partition.

No I can't say this with 100% certanty because I didn't do the testing, but ANYTIME you require more vram than a GPU has its performance has a huge penalty......which didn't happen to the 780Ti. Therefore the drop in the 970 performance relative to other cards is linked to another limitation within the cards design.

Take it for what its worth...I am done trying to talk to people about 970 limitations (yes some exist, but they aren't all linked to one flaw).
 
xentr_thread_starter
Truth of the matter is that it is impossible to determine what the cause is. However, it is shortsighted to arbitrarily blame the memory partition with the amount of evidence pointing to a contrary theory.
 
Sagath said:
This is exactly why it cannot be the memory bus actually. If the 780Ti stays strong with 4GB of textures loaded, then 1GB of textures are being cached off card at 22GB/s, instead of its memory bus coming in to play at 336gb/s. Whereas the 970 would have 3.5GB running at (196) and another 512 running at (28), and none at all running at 22GB/s cached off card.
Click to expand...
I'm thinking strictly bus width, not memory bandwidth. I have no technical explanation, only the observation that the one obvious similarity between the cards that gain and surpass the 970/980 all have larger bus widths.
 
SKYMTL said:
Truth of the matter is that it is impossible to determine what the cause is. However, it is shortsighted to arbitrarily blame the memory partition with the amount of evidence pointing to a contrary theory.
Click to expand...

If the 970 is so close to the 780ti and sometimes worse in performance, isn't it a safe bet that nvidia intended all along to release something in the short term that far surpasses the 970s and 980s we know today, and that they are merely the test bed/proving grounds for something far superior, but I guess that's the way it always goes isn't it.

I'm looking for justification for buying the 970 and reasons to stop thinking it's crippled, but I can't help but think it's going to hit it's performance ceiling sooner than later. I probably should stop reading https://forums.geforce.com/default/board/155/geforce-900-series/ as it's full of unhappy people. Did I tell you I got a reply from MSI for an 'upgrade offer' to a 980: $195USD if I return the 970. Not much of a savings when you factor in the regular prices of the two cards.

This is where my pain comes from: I buy hardware infrequently enough so that it lasts my needs longer than the average person, the gtx560ti lasted almost 4 yrs and still works. This 970 was the most expensive single piece of computer hardware I've ever bought in 15+yrs of pc building. Now I'm thinking I'll be unhappy with it in 6-12 months because no one could predict that Grand Theft Auto 5 or some other 2015-2016 game wouldn't run smooth on it with relatively high settings at 1080p. For those reasons I haven't written off the 980 upgrade offer completely, but not for $195US.

edit: just wanted to add that I know this is like beating a dead horse but I'm honestly not trying to fan the flames
 
Last edited:
xentr_thread_starter
There are a few things to take into account.

In many cases the card will likely be SM or TMU limited far before memory allocation steps in to throw down performance. As it stands, in the vast majority of situations, the GTX 970 is already operating at sub-par levels (IE: under 40FPS) before hitting the 3.5GB mark. I have yet to see any indication that 2015 and 2016 titles will require 3.5GB of memory at 1080P unless there are some serious issues with optimizations.

Another thing to take into account is DX12. While no one can predict how well it will penetrate over the next two years but it could have a drastic effect upon overall performance metrics.
 
You must log in or register to reply here.

Latest posts

xentr_legal_notice_title

xentr_legal_notice_description

Back
Top