WCCftech: Memory allocation problem with GTX 970 [UPDATE] PCPer: NVidia response

[AnandThenMan]

Not at all. Instead of some kind of rubbish press release, AMD are putting their money where their mouth is and offering something to entice nVidia users to jump ship.

From a marketing standpoint, Roy had perfect response, at the perfect time.

Agreed, it's low key and hits the target perfectly. Well done AMD (for a change).

IF that's the truth then they'll have to disable the last 512MB. There's no way to fix that.

There IS no way to fix the hardware oddity through software. The best Nvidia can hope to do is mitigate it, there is no cure. Honestly the best way is to just cap the memory at 3.5GB and be done with it. For non-gaming stuff the last 512meg block could be used seeing consistent memory speed is not so much a concern.

 

[Cloudfire777]

What? How did you arrive at that number?!

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;

looks like 50% performance hit.
2 seconds read performance:
Advertised:
224GB/s x2seconds = 448GB
Actual:
196GB/s x1second = 196GB
28GB/s x1 second = 28GB
Total of 2 seconds = 224GB

That is incorrect

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.

http://www.pcper.com/reviews/Graphi...Full-Memory-Structure-and-Limitations-GTX-970

 

[Head1985]

Computerbase analysis
http://www.computerbase.de/2015-01/geforce-gtx-970-vram-speicher-benchmarks/

 

[Erenhardt]

That is incorrect

http://www.pcper.com/reviews/Graphi...Full-Memory-Structure-and-Limitations-GTX-970

But its just theoretical.
Ryan said that in real life it aint gonna happen:

Despite all of this, achieving peak memory bandwidth performance on the GTX 970 is still possible, but it requires much more effort since simple striping will not do the trick. The easiest and most effective solution in this regard is to interleave reads and writes over the segments, such that one segment is writing while another segment is reading. Interleaving in this fashion allows both segments to work at once – avoiding the blocking effect of the shared read and write buses – and makes it more likely that both segments are doing useful work rather than waiting for their turn on an operation. However because this is only applicable to situations where more than 3.5GB of VRAM is in use and both segments are necessary, this means it's only theoretically possible to achieve 224GB/sec when more than 3.5GB of VRAM is in use. In any situations below 3.5GB we are de-facto limited to just the larger segment, in which case there are only 7 memory channels (196GB/sec) to stripe memory operations across. NVIDIA could of course interleave operations sooner than that and use both segments more often, but due to the blocking effect we've discussed before the performance hit from using the 512MB segment can quickly become greater than any gains.

 

[Enigmoid]

But its just theoretical.
Ryan said that in real life it aint gonna happen:

And again, you won't see that. Bandwidth to shader ratio is the same as the 980, if not better

 

[RussianSensation]

Well, I'm glad I got a 980 and not a 970. But I never even considered a 290. The 9x0 make a lot less noise regardless of how advanced cooling solution you get for the 290.

Your opinion contradicts real world user experience. You cannot say universally that all R9 290/290Xs produce more noise than any GTX970. For example the popular EVGA GTX970 is way louder than the Sapphire Tri-X 290:
http://forums.anandtech.com/showpost.php?p=37125547&postcount=34

While you can measure the difference in the noise levels with a very accurate scientific tool, can you ear actually perceive the difference in an enclosed case over your PSU, CPU and case fans? Can you hear a difference of 6-7dB in a Fractal Design R5? Don't forget that some heatsink fans make lower noise but it's more uncomfortable due to a certain electrical sounds/frequency of the ball bearings. It's not that simple.

See this video of how quiet after-market R9 290s actually sound under max load:
http://www.youtube.com/watch?v=pYxmW4JiJs8

It seems you also forgot another major fundamental aspect of balance with respect to heating and cooling vs. noise levels. If an ASIC is rated to run at XX*C, I can adjust my fan curve to reduce noise levels and still operate within the safe operating limits of the ASIC. Your post implies the noise levels on R9 290 cards are fixed at what you claim are loud levels.

Why can't a PC gamer customize a fan curve on say the MSI Lightnings to 85-90C considering they operate at 70-71C at 1.15Ghz overclock? It's not as if Hawaii will fail at 85-90C. If you care about noise levels that much, you can easily run them at 87-88C at even quieter noise levels.

http://www.tomshardware.com/reviews/r9-290x-lightning-performance-review,3782-7.html

Its low life pathetic marketing at best.

Baseless opinion. You call it pathetic marketing, I call it an opportunity to get some PC gamers informed. Too many PC gamers have only used NV for 5-10 years and wouldn't' even consider AMD due to blind brand loyalty. Try to lure in a customer with a lower cost product is a smart strategy. If an NV user doesn't like the AMD R9 290X card, it didn't cost him that much to try it out, considering the performance is so close to a 980. You make it sound as if switching to AMD is some doom and gloom outcome....
http://www.techpowerup.com/209412/amd-cashes-in-on-gtx-970-drama-cuts-r9-290x-price.html

Diamond R9 290X (rebadged HIS IceQ2 290x) after MIR = $280
Asus R9 290X after MIR + coupon = $280
Gigabyte Windforce R9 290X after MIR = $280
MSI Lighting R9 290X (the best 290X) = $320

vs.

Asus Strix 980 = $545
MSI Gaming 980 = $550
Gigabyte G1 980 = $610

I am sorry but even if CF doesn't work, one gets just 9-11% less performance at 1440P and up, but when CF does work, it's a ridiculous advantage for the 290Xs.

That's before even talking about even cheaper R9 290s. Unless all you do is play GW titles or run a miniITX, it sounds to me like the "pathetic marketing" being employed here is for the $550-600 GTX980. :thumbsdown: For any brand agnostic gamer, the comparison of a $640 MSI Lightning 290Xs vs. a $610 Gigabyte G1 980 is an eye-opener to how out of line the pricing disparity has gotten between NV and AMD's offering all because of perf/watt, power usage and brand loyalty. It's a really sad time for the GPU industry as there is no way something like this would have ever happened during ATI/NV days if ATI's best card was only 10% slower but cost HALF. :sneaky:

I never got into PC gaming because of power savings. If I did, I would have been on a Wii U or PS4 and never overclocked components. Even high-end gamers eventually realize when the value of the competition is simply awful.

To date, NV has not offered any monetary compensation, any discount towards a future NV GPU purchase, not even a game voucher. The sad part is so many 970 owners will probably either step down to a 960 or get a 980 instead of trying AMD because they heard from some guy out there that ATI/AMD drivers suck from 10 years ago.

 

[Silverforce11]

Computerbase analysis
http://www.computerbase.de/2015-01/geforce-gtx-970-vram-speicher-benchmarks/

This shows how terrible a lot of the USA review sites are, particularly TR, PCPER & Hardware Canucks, they have shown themselves to be utter NV-shills. When NV told them to jump on the FCAT & smoothness bandwagon to bash AMD, they did so in a hurry. When NV's Maxwell suffer poor smoothness, they stop using FCAT or focus on frame time.

Now this problem is revealing itself to be very ugly when people with a clue examine frame times.. these shill sites do nothing, a whole bunch of AVERAGE FPS benches.

This is just an embarrassment:

The 970 is gimped. That is unacceptable stutter, for a single GPU. In SLI, the problem is magnified, and is ridiculously bad. Anyone who thinks that's not a problem is an utterly delusional.

http://hardforum.com/showpost.php?p=1041392263&postcount=619

 

[96Firebird]

This shows how terrible a lot of the USA review sites are, particularly TR, PCPER & Hardware Canucks, they have shown themselves to be utter NV-shills.

Lol...

 

[Silverforce11]

Lol...

Would it change it if I said North America? Point still stands. Lol...

 

[AnandThenMan]

I'm finding PCPer especially questionable.

 

[flash-gordon]

The Talos Principle absurd frame time shows what to expect from this card on games not specially optimized or when the driver support slows down in the future...

 

[Silverforce11]

Video proof:
http://www.computerbase.de/2015-01/geforce-gtx-970-vram-speicher-benchmarks/4/

What's worse is Computerbase.de is saying if you have a 2nd monitor attached showing the desktop (not multi-monitor res, just single monitor gaming), it eats ~200MB of vram, so the 970 suffers a lot more stutters than it normally does.

http://www.computerbase.de/2015-01/geforce-gtx-970-vram-speicher-benchmarks/5/

Not only that, the problem is worse in SLI, because SLI reserves a portion of the memory to function so the 3.5GB limit is lowered thus users will hit that break-point earlier.

Now wonder users reported horrific stutters in SLI.

 

[tential]

Video proof:
http://www.computerbase.de/2015-01/geforce-gtx-970-vram-speicher-benchmarks/4/

What's worse is Computerbase.de is saying if you have a 2nd monitor attached showing the desktop (not multi-monitor res, just single monitor gaming), it eats ~200MB of vram, so the 970 suffers a lot more stutters than it normally does.

http://www.computerbase.de/2015-01/geforce-gtx-970-vram-speicher-benchmarks/5/

Not only that, the problem is worse in SLI, because SLI reserves a portion of the memory to function so the 3.5GB limit is lowered thus users will hit that break-point earlier.

Now wonder users reported horrific stutters in SLI.

The tests are one thing. But what people need to realize is that it doesn't replicate real world usage at times.
Which is why I think the GTX 970 memory issue is larger than it appears from just testing alone. I think there are a lot of scenarios where people will see issues that the "tests" didn't show because it's not a true "4GB" card.

People are focusing at a very narrow window of current releases which obviously Nvidia has already optimized for rather than the big picture.

 

[Silverforce11]

The tests are one thing. But what people need to realize is that it doesn't replicate real world usage at times.
Which is why I think the GTX 970 memory issue is larger than it appears from just testing alone. I think there are a lot of scenarios where people will see issues that the "tests" didn't show because it's not a true "4GB" card.

People are focusing at a very narrow window of current releases which obviously Nvidia has already optimized for rather than the big picture.

Umm, it does. It's playing games at 1440/1600p.

They noted the problem is worse for SLI setups.

These are GameWorks games, obviously should already be optimized for Maxwell, yet it still stutters and drop frames like crazy on the 970 due to the 3.5gb vram segmentation.

 

[Carfax83]

I think you have to look at those results in context, to get an accurate perspective. It seems to me that a lot of these reviews are forcing the GTX 970 to play at settings that it's simply incapable of doing.

I have AC Unity, and I'll tell you that it is not playable on GTX 970 SLI at 1600p with 4x MSAA, which is exactly how computerbase.de tested it. Also, the engine itself can cause stuttering due to it's inefficiency, which they do mention in the article.

The same goes for Far Cry 4. The GTX 970 does not have the bandwidth for high levels of MSAA in graphics heavy games that use deferred rendering..

 

[Silverforce11]

That's single GPU testing, 30-40 fps. SLI should be able to handle that smoothly, if it had the vram. Note the problem does NOT occur for the 980 which is allowed to freely use its 4gb vram buffer.

 

[Carfax83]

Umm, it does. It's playing games at 1440/1600p.

They noted the problem is worse for SLI setups.

These are GameWorks games, obviously should already be optimized for Maxwell, yet it still stutters and drop frames like crazy on the 970 due to the 3.5gb vram segmentation.

You didn't even read the review properly. They didn't test 1440p anywhere. The lowest resolution they tested was 1600p, and that was with high levels of MSAA. Then they also tested at 1960p, 4K and even beyond 4K.

People forget these cards have a 256 bit bus, and will choke at such high resolutions when you add 4x MSAA to the mix.

 

[AnandThenMan]

Note the problem does NOT occur for the 980 which is allowed to freely use its 4gb vram buffer.

Exactly. And I have no idea why playing the game is somehow an invalid test.

People forget these cards have a 256 bit bus, and will choke at such high resolutions when you add 4x MSAA to the mix.

The 970 does not have a 256 bit bus.

 

[Silverforce11]

You can see the problem at 1600p. Users report the problem at 1080p (Shadow of Mordor, Dying Light) and 1440p (These same games Computerbase.de tested) also.

The 980 (256bit bus!) under the same scenario doesn't choke up because it actually has full speed 4gb and not crippled dual partitions.

So we've gone from "there's no problems" to "don't run at high settings"...

 

[cmdrdredd]

I think you have to look at those results in context, to get an accurate perspective. It seems to me that a lot of these reviews are forcing the GTX 970 to play at settings that it's simply incapable of doing.

I have AC Unity, and I'll tell you that it is not playable on GTX 970 SLI at 1600p with 4x MSAA, which is exactly how computerbase.de tested it. Also, the engine itself can cause stuttering due to it's inefficiency, which they do mention in the article.

The same goes for Far Cry 4. The GTX 970 does not have the bandwidth for high levels of MSAA in graphics heavy games that use deferred rendering..

I can attest to that. I can't use 4x MSAA at 2560x1440 in FarCry 4 but I can run at 2804x1577 and it works fine with minimums around 50fps.

 

[wanderer27]

The more I read about this, the more I thing they'll effectively have a solution for single GTX 970 users.

SLI is where the real negative impact is going to remain, though that should be alleviated somewhat.


My thoughts on the solution is that they can dedicate the upper 512MB to OS/Desktop/low demand graphics which should really show no impact.

Now I've seen reports that this "low speed/demand" graphics can require anywhere from 200MB to 500+MB (varies).

Optimum case being 500MB, then that would leave the 3.5GB for fast access.
Compared to 980 usage, you'd still have 3.5GB for fast access, but the 500MB would also use fast access.

Essentially in this case it's a wash . . . well, actually 970 would get an improvement >3.5GB over what it is now since the slow Memory segment is locked to the "low speed/demand" graphics.


Now SLI best case, you could possibly split the "low speed/demand" graphics between the two cards, so in this instance you could gain 250MB of high speed access.

Should still see an improvement in higher demand performance, just not as much as a single 970 (which should essentially be running at full speed now).




.

 

[Abwx]

It seems to me that a lot of these reviews are forcing the GTX 970 to play at settings that it's simply incapable of doing.

Even the 970m is advertised as 4K on MSI site, i guess that they didnt decide alone that it was 4K "compatible"...

Résolution 4K

http://fr.msi.com/product/nb/GS60-2QE-Ghost-Pro-4K-Gold-Edition.html

MSI GeForce GTX 970

Supports 4K Displays

http://www.rebeltech.co.za/8899-msi...ng-n970-gold-gold-edition-4gb-256bit-ddr.html

 

[AnandThenMan]

Nvidia on their site point to at least some of these reviews, all of which have the incorrect hardware specs (which were supplied by Nvidia). So even though NV has admitted the given specs were wrong, people reading reviews will still get the wrong info. I find this quite unbelievable.

http://www.tomshardware.com/reviews/nvidia-geforce-gtx-980-970-maxwell,3941.html
http://techreport.com/review/27067/nvidia-geforce-gtx-980-and-970-graphics-cards-reviewed
http://www.pcgamer.com/gigabyte-geforce-gtx-970-g1-gaming-review/
http://hothardware.com/Reviews/NVIDIA-GeForce-GTX-980--970-Maxwell-GPU-Reviews/
http://hardocp.com/article/2014/09/29/msi_geforce_gtx_970_gaming_4g_video_card_review/1

 

[Silverforce11]

The more I read about this, the more I thing they'll effectively have a solution for single GTX 970 users.

SLI is where the real negative impact is going to remain, though that should be alleviated somewhat.


My thoughts on the solution is that they can dedicate the upper 512MB to OS/Desktop/low demand graphics which should really show no impact.

Now I've seen reports that this "low speed/demand" graphics can require anywhere from 200MB to 500+MB (varies).

Optimum case being 500MB, then that would leave the 3.5GB for fast access.
Compared to 980 usage, you'd still have 3.5GB for fast access, but the 500MB would also use fast access.

Essentially in this case it's a wash . . . well, actually 970 would get an improvement >3.5GB over what it is now since the slow Memory segment is locked to the "low speed/demand" graphics.


Now SLI best case, you could possibly split the "low speed/demand" graphics between the two cards, so in this instance you could gain 250MB of high speed access.

Should still see an improvement in higher demand performance, just not as much as a single 970 (which should essentially be running at full speed now).
.

So you think just setting it as 3.5gb would alleviate the problem? Why didn't they just launch it fused off or via bios to only have 3.5gb accessible in the first place, it would not be an issue to sell 970s for its price with that vram. This way users know whats on the table and can make a more informed choice.

I mean if NV told you its 3.5gb, you see a clear difference to the 980. So you wouldn't be surprised when you can't handle ultra textures or MSAA as well. That would make some users pay extra for the 980s!

 

[Magic Carpet]

The tests are one thing. But what people need to realize is that it doesn't replicate real world usage at times.
Which is why I think the GTX 970 memory issue is larger than it appears from just testing alone. I think there are a lot of scenarios where people will see issues that the "tests" didn't show because it's not a true "4GB" card.

People are focusing at a very narrow window of current releases which obviously Nvidia has already optimized for rather than the big picture.

As a multi-monitor user, I can second that. You need extra VRAM for even blank screens.

People forget these cards have a 256 bit bus, and will choke at such high resolutions when you add 4x MSAA to the mix.

Well, the people are not able to test it with a flagship card. Instead, they are limited to mid-range cards at flagship prices... Yeah, you are right. We are expecting way too much. Hopefully, AMD will be able to fix that with a proper card.