The Real Nvidia GeForce GTX 970 Specifications

Page 5 - Seeking answers? Join the Tom's Hardware community: where nearly two million members share solutions and discuss the latest tech.
Status
Not open for further replies.


The engineers knew but the PR department was confused which is perfectly understandable as I couldn't grasp why you'd partition memory at first.

It's quite common to have miscommunication like this in a big company. The real issue is whether there's anything to be concerned about if a game uses between 3.5GB and 4GB and if true whether NVidia can provide a fix.
 


These new cards use lossless compression so need less bus width. You can't simply compare a similar AMD or older NVidia card due to this.

(I'm not disagreeing with anything above but just pointing this out)

This is why I tell people to simply look at the game benchmarks and don't get too hung up on engineering details (except for perhaps the memory partition issue which I'm still investigating to see if stutter is possible in the 3.5GB to 4GB range).
 
Update:

I investigated further and discovered that the ONLY real issue is the memory partitioning. Furthermore all reputable sources are reporting that this at worst reduces the frame rate by 3% over what it would have been with a full ROP count and full speed access to all 4GB.

3%!

So people are whining about only having 3.5GB (not quite true) and having their games suddenly start stuttering if using between 3.5GB and 4GB.

*On the stutter issue (which isn't confirmed but MAY be true):
"Nvidia clamied this would be around a three percent drop. It's worth noting that Nvidia's benchmarks only looked at average FPS performance, which may not account for the frame stuttering that some users claim to be experiencing."

So it's still a little confusing.

I like NVidia but I'm not a fanboy. I'm not too concerned with this but I will be looking closer and won't get a GTX970 unless there's a fairly simple fix (if it's even an issue).

The last 512MB gets treated like a victim cache under normal operation as it's still faster then system memory. The reason the stuttering isn't verifiable is that you have to force the system to load new data into the last 512MB which then forces it to read from that region during a frame draw, it's basically bypassing the drivers optimization. Not many people do that.

The excuse that they didn't know the inner workings of a product they were advertising is scary.

Not an excuse, it's how company's run. You have different departments and each department has different specializations. The technical folks that do R&D know all about this design, they are the ones that engineered it as a work around to the fact that one ROP will be defective on some GM204 chips and instead of throwing them all away they disable the defective area's and sell them to OEM's to produce 970's. This is standard industry practice and all your CPUs, APUs and GPU's use this technique to lower production costs and provide a full range of products. The guys in marketing only understand how to sell and advertise things, they aren't technical and most won't play video games or know the first thing about "graphics cards". Instead the technical guys do a white paper describing their product with a technical summary that highlights the important specifications. They present this paper to the marketing folks to do the brief and put together the advertisements and review kit. It's the job of the marketing folks to transcribe and paraphrase the technical summary and create the material that gets into everyone elses hands.

That is where the screw up was at (according to Nvidia), someone in marketing didn't transcribe the technical summary properly and listed the specifications incorrectly for the 970. This incorrect information got put into the review kit and pushed out the door. And since only a very small minority of people actually read the technical whitepapers and instead just rely on the marketing material, everyone had the wrong info during review and release. And because the difference is so small, nobody caught it (according to NVidia) for a very long time.

Personally I'm apt to believe them, having worked in corporate I frequently deal with the kinds of miscommunications that this situation relies on. This kind of stuff happens far more often then people know. Most people don't care because it's still the same product, it's not a bait and switch, just a misprint of technical information that's usually corrected via addendum.
 


this is a bit more on the hearsay side of things, though from some people with both 980s and 970s on another forum apparently it's a very visible phenomenon if you try to create the scenario. it seems like there is quite a bit of micro stuttering that occurs when you force the card to use full 4gb of vram; apparently even single cards will look like bad xfire cards in this situation.

Reviewers probably didn't notice it mostly because 4k is the only resolution you can reliably hit 4gb of vram usage on a "normal" gaming benchmark and FCAT doesn't work on 4k, furthermore the framerates at 4k are low anyway, causing reviewers to turn down settings like AA which incidentally results in less vram being used, and even if it still was using 4gb of vram i'm sure the reviewers were writing off the microstutter to the less then 60fps results from the benching.

that said if you play a heavily modded game on a monitor that can keep up (say heavily modded skyrim at 1440p /144hz), apparently it can microstutter badly while kicking out 100+fps, i'll have to go back through some hundred page forum posts on another board to pull up some other examples but this certainly does seem to cause a degree of performance loss/microstutter
 
this is a bit more on the hearsay side of things, though from some people with both 980s and 970s on another forum apparently it's a very visible phenomenon if you try to create the scenario. it seems like there is quite a bit of micro stuttering that occurs when you force the card to use full 4gb of vram; apparently even single cards will look like bad xfire cards in this situation.

It's not just more then 3.5GB for a game, but more then 3.5GB for a single frame. That 512MB region is active, the driver is using it internally as a victim cache for older graphics data once the 3.5GB is getting full. The reason it's so rare is that you really have to screw around with settings to cause it to happen, basically find a way to override the drivers internal optimizations, or just turn everything up so high that your doing over 3.5GB per frame, which is usually ridiculously. We're talking 4K, 16X AA, Ultra settings here which would have this card going unplayable fps anyway. Keep going and you end up with more then 4GB of per-frame data which would make everyone start hitting the system memory pool and cause all sorts of hell.
 


I agree

The drivers were designed not to use that last 500mb of vram whenever and where ever possible, though i've read there are some situations people have been able to corner the card into using it.
 
You failed on your car analogy - more like it performs perfectly up to just about redline - THEN IT DOES NOT!!!! There is no way to "spin" this performance issue after the 3.5G boundary is reached. I bought this card, like many others, to keep for a while and possibly SLI & upgrade my primary gaming monitor. This option has been taken away because this card has flaws(not features) that were falsely reported by nvidia. I have owned an 8800, GTX 460 1GB SLI, GTX 750Ti, GTX 670 and now a GTX 970. As you can see, I am a loyal nvidia consumer - my faith has been shaken.
-Bruce
 

SBMfromLA

Distinguished
It makes me wonder what would have happened had this card been originally marketed as a 3.5GB + 512MB Reserve card? It seems like everyone's perception might have been altered into thinking that setup was awesome.
 


that's possible, still i think the choice between the r9-290x and gtx970 would become much more muddled then it was when it was announced as 4gb. the main consideration for 4k "futureproofing" is large amounts of vram. when everything was equal you had two cards with the same vram and roughly the same performance only one drew 1/2 the power of the other. it seemed like a no brainer.

Now if you're planning of future proofing or building for 4k, the reality is the r9-290x becomes a more attractive card then the 3.5gb 970. I feel for the people who bought one thinking they were getting the better card for this scenario and understand their upset at realizing they made their purchase decision with bad info. Heck i nearly replaced my r9-280x with a gtx970.

glad i held off for a bit, i feel like a dodged a bullet.
 


yes you did buddy. I ditched my 7970 for a gtx 970 and wish I stayed in the red camp :(

to make matters worse I got a waterblock for my 970 so I can't exactly return it
 
Similar to the gtx660 with 2gb 192 bit memory. The card never used more than 1.5gb, as onlt 1.5gb of it is 192bit and the remaining memory only had access to a 64bit link. I sold both of mine because it WAS causing issues in some games, as the driver or firmware or something wouldnt allow the card to use that slower memory, so it was behaving as a 1.5gb card.
 
i have created the scenerio with my 3gb 780. run uncompressed 8/16k texture packs and watch your fps go to under 10.

this is ridicoulous.

i understand that saying its 4gb but, kinda not really, only 3.5gb fast, other 0.5gb slow.... whatever.

but nobody had vram problems on 3gb 780s or 7950s/7970s/refresh/refresh, and now all of a sudden... 970 users are attempting to use settings that will fully crash out a 780 one it passes 3gb, they still have 0.5gb of headroom until 3.5gb is reached, and then complain. for all intent purpose the 970 is roughly equal in performance to the 780 in games. so how all of a sudden are people trying to shove settings in that even a 780ti cant hang with then complain?
 


how this was discovered is a bit strange... but at this point it's a marketing issue. Nvidia's marketing team messed up in the way they listed the specs... and now they're trying to shove it under the rug, which is why people are mad
 
for all intent purpose the 970 is roughly equal in performance to the 780 in games.

No it's not. Spend time and read the details about the different configurations. If there are concepts that you are unfamiliar with, don't be afraid to ask questions from others more knowledgeable or even do your own deep research about them. The 970 has seven fully active 32-bit GDDR5 memory channels that are interleaved across 3.5GB of GDDR5 memory. This produces a bandwidth of 224-bits during most scenarios. The last 32-bit GDDR5 memory channel is slaved to the seventh as it's L2 cache and ROP have been disabled, it has it's own 32-bit GDDR5 connection to a 512MB GDDR5 DRAM chip. The 980 has eight fully active 32-bit GDDR5 memory channels that are interleaved across 4.0 GB of GDDR5 memory for a full bandwidth of 256-bit. These cards, while similar, are definitely not the same performance.
 

jdw_swb

Distinguished
Feb 11, 2008
368
0
18,810
I suppose no-one from Nvidia read any reviews of the GTX 970 from the major tech sites?

No-one at Nvidia noticed that the specs listed on the front page of most reviews were different than what was actually true?

Impossible, they even link to these reviews on your Official GTX 970 page.

http://www.geforce.com/hardware/desktop-gpus/geforce-gtx-970/reviews

I don't believe that they didn't know about the mistake before. After the first reviews went live, Nvidia should have contacted the sites immediately to correct the specifications error.
 
are 3gb gk110 users feeling the same issues that 970 users are feeling?

i dont understand how a 970 can run any higher settings than 780ti users are running.... yet, we havent heard about this vram issue yet. its been almost 2 years now and we have had less vram just as powerful card.

i call bs on the performance issues.
 


I'm using two 780 Hydro Coppers in my rig and I have yet to experience any issues running 1440p on an ASUS ROG Swift. I also don't go out of my way to try to force the drivers either. There are very few situations where someone will be using more then 3GB of graphics memory, and when those situations do occur a 970, even a 980, would be stuttering along at unplayable slideshow speeds. Someone has to really enjoy playing at under 30fps to see any problem. We're talking 4K Ultra settings here, which not even a 980 can do. People trying to argue "but future games!!!" are forgetting that memory usage is largely dependent on AA levels and both cards are already struggling at these supposed "future" resolutions. Tomorrow game isn't going to play any less then todays games in this regard.
 


you are misunderstanding the problem. this isn't about having x amount of vram isn't enough, type of thing. If this card was a 3gb card or a 3.5gb no one would complain. However it was SOLD as a 4gb card with a 224bit bus.

- It doesn't have a 224bit bus, it has a 196bit bus for the first 3.5gb of ram. + 28bit bus for the last 500mb of ram
- it doesn't have 4gb of gddr5 ram it has 3.5gb of gddr5, and 500mb of gddr5 ram that runs at ddr3 speeds (or slower)
- it doesn't USE 4gb of gddr5 ram unless it absolutely has to, then the whole memory structure slows down to roughly ddr3 memory speeds (not just the last 500mb) in order to balance out the memory interactions with the gpu. This causes a DRASTIC reduction in the performance of the gpu, micros stuttering being the most obvious example of what happens.
- it doesn't have the advertised 64 ROPs, it has 56
- it doesn't have 2mb of l2 cache it has 1.75

this is not the card it's claimed to be on the box. and if you're in a situation that calls for 3.5gb of vram or more, your Ferrari will rapidly turn into a pinto. Remember people chose this card over a r9-290x, a card which has a FULL 4gb of vram, primarily because it drew half the power under load, and performed pretty much identically. Well guess what. in heavy vram using games it won't perform identically, it will perform worse.

which is why people are upset. they bought a card with the best information they had available at the time, thinking it was the best card for their needs, and in the end it turns out the r9-290x probably would have met their needs better (mostly, depends on why they chose the 970. there certainly are buyers of the 970 who are not and never will be affected by this, that said many of the reasons for buying this card over the similarly priced 290x are invalidated by these actual specs)
 
- It doesn't have a 224bit bus, it has a 196bit bus for the first 3.5gb of ram. + 28bit bus for the last 500mb of ram
- it doesn't have 4gb of gddr5 ram it has 3.5gb of gddr5, and 500mb of gddr5 ram that runs at ddr3 speeds (or slower)
- it doesn't USE 4gb of gddr5 ram unless it absolutely has to, then the whole memory structure slows down to roughly ddr3 memory speeds (not just the last 500mb) in order to balance out the memory interactions with the gpu. This causes a DRASTIC reduction in the performance of the gpu, micros stuttering being the most obvious example of what happens.

Horrifically incorrect despite many attempts to state the correct info.

It has seven 32-bit GDDR5 memory controllers for the first 3.5GB, that is 224-bit memory bus. The eighth 32-bit controller is slaved to the seventh and provides a single 32-bit bus. All chips run at GDDR5 speeds, there is no DDR3 inside this unit. The total combined memory bandwidth is indeed 256-bit at full GDDR5 speeds, problem is that since the eighth ROP/L2 unit has been removed it can't control all eight memory controllers at once for the vast majority of work loads. The seventh controller can only issue one read and one write per cycle. What happens is the first seven controllers are issued a read command, then after it returns the seventh is issued the read command for the eighth. All channels have the exact same bandwidth so the result is doubling your read/write latency. For this reason you don't want to go over 3.5GB of graphics storage and let the card use the last 512MB as a victim cache.

There is exactly 4GB of GDDR5 memory inside this unit, no more and no less.

It most certainly does use all 4GB of GDDR5. The driver will utilize the last 512MB as a form of L3 victim cache to hold graphics data before dumping it into system ram. This is invisible to the user, just like all caching mechanisms. You don't get to see what's inside your L1, L2 and L3 caches for system memory, and neither do you get to see inside this GPU's cache. What you can do is override the drivers attempts at cacheing and force it to use this memory as graphics RAM, which will present a performance penalty as you just doubled your VRAM's read latency whenever you attempt to read from a region that includes that eighth channel.

There is a 3~5% performance penalty, as demonstrated by many independent sites. There is no "DRASTIC" penalty and the micro-stuttering is minor at worst.

Now I'm going to copy past a previous post I made to another person spreading the same misinformation you are with the same warning and consequences.

How the heck are you not reading? Toms just did a break down of it, Anands did a really good technical breakdown describing exactly why everything is the way it is. And still you put out false and misleading information. I'm going to spell it out to you now so that you can't use ignorance as a defense anymore.

The GM204 chip has eight 32-bit GDDR5 memory channels. Each memory channel is connected to a ROP unit that contains 256KB of L2 cache and eight individual ROPs. Previously if there was a defect in one of those eight ROP's or in the associated L2 cache memory, the entire component would have to be disabled which also disables the associated 32-bit GDDR5 memory channel. Maxwell has the ability to selectively disable individual subcomponents, so instead nVidia only disables the L2 cache and connected eight ROP's but leaves the memory channel operational. Now since there is no cache or connection to the crossbar, that memory channel is then slaved to a nearby fully operational unit. Memory data is typically interleaved across all channels in 1KB blocks, so if you needed to read 4KB of data it would exist on four different 32-bit GDDR5 chips and would be read at the speed of x4 of a single chip. In the case of a fully operational GM204 chip, the 980, the entire eight 32-bit GDDR5 channels are used for an eight way interleave. For binned GM204 chips, the 970, only the fully independently operational seven channels are used for a seven way interleave. That last slaved 32-bit GDDR5 chip is kept separate as to not lower the performance of the rest and instead used as a victim cache since it's still faster then system memory. That is how and why you got your 3.5GB / 512MB segments, and why putting them all into a single 4GB eight way interleave segment would be disastrous for performance (two of the eight interleaves couldn't talk at the same time so every stripped access would have double the latency).

So ultimately what you are complaining about is why didn't nVidia give you a fully operational 980 when you bought a binned 970. Yes there was incorrect marketing material, and that is something that nVidia is going to have to deal with internally, I expect some middle manager is going to be asked to fall on his sword for this.

Now anymore deliberate spreading of incorrect technical information will be met with a vacation. You can have whatever personal opinion you want for whatever reason you want, just don't go around telling others lies.
 

spellbinder2050

Distinguished
Sep 7, 2008
175
0
18,680
"We can empathize with buyers who feel betrayed, though. Nvidia definitely has some mind-share to earn back. But to us the price/performance ratio trumps everything else, and that is no different today than it has been since the GeForce GTX 970 was released."

price/performance trumps the accuracy of the product description, especially given that there are competing gpu makers? It trumps knowing whether your card will be future proof? Not really.

Personally I feel betrayed. Nvidia better make this right somehow.
 

rdc85

Honorable


could u copy the link here..

I'm got one guru3D test that claim that..
even i saw some flaw in his reasoning
the 970 vram it used to compare is 3.5 Gb vs 3.6 Gb
and not 3.5Gb vs 3.8 Gb - 4 Gb some forum user reporting..

(3.6 is above 3.5 but the 0.1 maybe just unnecessary data or the problem will worse if u increased it to 0.2 0.3 0.4)
 

xHopesFall

Distinguished
Feb 6, 2011
19
0
18,520
I really don't understand what people are bitching about... It performs just as well as it did before you knew the 'Numbers'... Go cry about semantics somewhere else people... Please don't turn TH into facebook rants, and reddit kid crying sessions. Performance in the card changed none.

If you Were coloring with a blue crayon, and it looked blue. Turns out the wrapper says green... are you really going to cry about the wrapper being green when you wanted blue?.. IT IS BLUE!
 


Guru3D, PCPer and Hardwarecanuks

http://www.pcper.com/reviews/Graphics-Cards/Frame-Rating-Looking-GTX-970-Memory-Performance

Closing Thoughts

I spent nearly the entirety of two days testing the GeForce GTX 970 and trying to replicate some of the consumer complaints centered around the memory issue we discussed all week. I would say my results are more open ended than I expected. In both BF4 and in CoD: Advanced Warfare I was able to find performance settings that indicated the GTX 970 was more apt to stutter than the GTX 980. In both cases, the in-game settings were exceptionally high, going in the sub-25 FPS range and those just aren't realistic. A PC gamer isn't going to run at those frame rates on purpose and thus I can't quite convince myself to get upset about it.

http://www.guru3d.com/news-story/middle-earth-shadow-of-mordor-geforce-gtx-970-vram-stress-test.html

Concluding

Our product reviews in the past few months and its conclusion are not any different opposed to everything that has happened in the past few days, the product still performans similar to what we have shown you as hey .. it is in fact the same product. The clusterfuck that Nvidia dropped here is simple, they have not informed the media or their customers about the memory partitioning and the challenges they face. Overall you will have a hard time pushing any card over 3.5 GB of graphics memory usage with any game unless you do some freaky stuff. The ones that do pass 3.5 GB mostly are poor console ports or situations where you game in Ultra HD or DSR Ultra HD rendering. In that situation I cannot guarantee that your overall experience will be trouble free, however we have a hard time detecting and replicating the stuttering issues some people have mentioned.

http://www.hardwarecanucks.com/forum/hardware-canucks-reviews/68595-gtx-970s-memory-explained-tested-2.html

They all came to the same conclusion, that NVidia's drivers were doing a great job of dealing with the memory configuration and there is very little performance penalty from going over 3.5GB. Remember the drivers are deliberately trying to keep new data inside the 3.5GB segment while evicting older data into the 512MB segment so a game actually needs a working set over 3.5GB instead of just storing over 3.5GB of graphics data into memory. This is a very similar situation to what your CPU is doing, there isn't 2GB of cache (super fast local memory) on your CPU, only 4~8MB, yet programs frequently are several times that size. The cache logic use's hardware to actively scan instructions and preload / data into the fast local cache before you need it.

This card should of been marketed more accurately as a 3.5GB GDDR5 VRAM + 512MB GDDR5 L3 cache since that is what it's doing. The only way you get any performance penalty is when you actively try to circumvent the cards logic and then you end up at unplayable frame rates anyway (25~30FPS on a 980). The only people who could possibly have issues are the folks who bough two for SLI and thus can get it to a semi-playable 40~48 FPS before they start to experience that 3~5% performance penalty. And at that point they can just go inside and turn the 16xAA down to 8xAA and problem solved. If they stamp and adamantly refuse to use a slight compromise on quality, then they shouldn't of been using lower shelf binned hardware (970) and instead forked over the money for perfect GM204 chips (980).
 


He was comparing the GTX 970 to the GTX 780, not the GTX 980.
 

CorbNoir

Reputable
Jan 28, 2015
2
0
4,510
I really don't understand what people are bitching about... It performs just as well as it did before you knew the 'Numbers'... Go cry about semantics somewhere else people... Please don't turn TH into facebook rants, and reddit kid crying sessions. Performance in the card changed none.

If you Were coloring with a blue crayon, and it looked blue. Turns out the wrapper says green... are you really going to cry about the wrapper being green when you wanted blue?.. IT IS BLUE!

Gee I don't know, maybe some of us were thinking more than a month into the future when, sooner rather than later, games are going to be properly taking advantage of 4GB and the card they bought under that (false) assumption isn't up to the task? It's not about how it performs now, it's about how this has a significant impact on the long-term usefulness of the card.

What's the deal with people who get so snide about others acting childish always beubg the ones with the most childish responses? Save your ad hominem garbage, it does nothing for you.
 
Status
Not open for further replies.