"It Won't Happen Again:" NVIDIA CEO Breaks Silence on GTX 970 Controversy

[HumanSmoke]

Obviously you don't remember that AMD got crap over it because it never specified a base clock and used the words "Up To".

Probably comes down more to degree ( 700+ posts in four front page articles and a few forum threads - something AMD's throttling didn't even approach even though it affected more than one SKU, and many user scenarios). Everyone kind of gets it as this point, and while it is fun for some people to pile on because the company is - lets face it, a stereotype tech villain, most of these threads are populated by people who don't own the card, and wouldn't buy Nvidia products in general. Great for blowing off steam, but in the greater scheme of things means close to nothing.

When there are twice as many forum posts about (un)/locking overclocking of Nvidia mobile graphics cards than there is about the NSA secreting spyware in hard drive firmware across the whole planet, it's probably time to take stock of what kind of yardstick you measure wrongdoing by...IMO

 

[nothappy]

When you receive an item you bought, you check it physically and tried your hardest to ensure of its quality. Reviewers do an even thorough job by putting it through the loops of benchmarks. But to truly understand it you need the engineers to determine it to the bone.

You can debate a 3.5 + 0.5 GB (of a lower spec) equals 4GB all you like. But as we know a Hyperthreaded 4 core is actually 2 REAL cores, and its not the same as a REAL 4 Core Chip. *nt*l said it out loud at the beginning, we bought it and loved it. If you think its not an apple to apple comparisson, its yours to decide.

So this letter is an undoubtable evidence of their screw up, and we have been given the right to do whatever we want with the 970, own your mistakes Nvidia, and shove it up your ...

 

[CrAsHnBuRnXp]

I believe the appropriate phrase is "Working as intended".....

This isnt Blizzard.

 

[Prima.Vera]

”...GTX 970 is a 4GB card. We wanted GTX 970 to have 4GB of memory...”
next time you should add memory card slot, an microSD card slot, ...up to 128GB memory...???!!!

Same issue where with the GTX660Ti graphics cards... same here...

Yeah, like in the good ol' days.
Remember S3 was doing that:

 

[heydan83]

So who will be buy? AMD? You remember when they released a card marketed at 1000MHz and then after a reasonable gaming session was found to drop down as low as 550MHz? It could last long enough to get through benchmarks, so the reviews were good, but actual user experience wasn't as good. Remember that? No? How about the time AMD released a card, let it go through all the reviews(which gave it very good reviews) and then released a driver that purposely and drastically reduced performance to stop the cards from dying pre-maturely? No. Don't remember that? Remember how they didn't even bother to apologize for either? Yeah, good times!

All you people need to get over your hate on nVidia because this is nothing compared to what some other companies have done and never even bothered to apologize for. What nVidia did was a marketing mistake, it doesn't affect the performance of the card we bought compared to the reviews we all read.

So you´re not willing to buy amd because of doing the same thing nvidia has done... then that doesn´t make any sense, I think we should make notice us as a good and power consumers and just skip this generation to let them know we have the control and not them...

 

[Naito]

That is the thing, having the 512MB only improves performance. Turning if off means that the card would start paging out to system RAM instead of accessing the 512MB, that is way worse.

Paging out to RAM will increase access times for non-critical stuff, but it also, to the best of my knowledge, does not create memory contention. If I understand correctly, a XOR situation occurs when one or the other segment needs to be written or read from, it causes a scenario where the other has to wait. This means any request to the 512MB segment has the possibility of harming the performance of the main 3.5GB segment.

EDIT: Either way, the GTX 970 will be a more driver dependent card, as Nvidia will probably have to spend more time working on how to manage this memory architecture with certain games.

 

[Xzibit]

Instead of being excited that we invented a way to increase memory of the GTX 970 from 3GB to 4GB, some were disappointed that we didn't better describe the segmented nature of the architecture for that last 1GB of memory.
This is understandable. But, let me be clear: Our only intention was to create the best GPU for you. We wanted GTX 970 to have 4GB of memory, as games are using more memory than ever.

How long until Nvidia tries to patent this invented way of increasing memory. Lets disable L2 cache re-route the controller and slow it down. BAM!!! we invented more memory.

He is also saying a 3GB card wasn't enough since more games are using more and more memory. Kind of shooting his explanation in the foot. You needed more memory but you slowed it down. Then have software to allocate less access data in the 512mb segment That sounds more like software/driver efficiency "optimization" not architecturally limited to Maxwell.

 

[btarunr]

That's the BS part. The driver team knew since they were allocating low access resources into the 512mb sector. It doesn't fly that there was a communication problem if 2 separate departments hardware and software knew with software developing drivers continuing to optimize for it. Reeks of deception or cover your ass because the lawsuits are coming.

The miscommunication was between engineering and technical-marketing (the team that writes the Reviewer's Guide). Software is another department.

 

[HalfAHertz]

Some of you are disappointed that we didn't clearly describe the segmented memory of GeForce GTX 970 when we launched it.

No, they're disappointing because they unknowingly got a cut down part...

We invented a new memory architecture in Maxwell. This new capability was created so that reduced-configurations of Maxwell can have a larger framebuffer

No, it was "invented" so that you could stick more RAM to crippled silicone and sell it for more $$...

This is a good design because we were able to add an additional 1GB for GTX 970 and our software engineers can keep less frequently used data in the 512MB segment.

If it was such an amazing design decision, why did you hide if from both your customers and the aforementioned software engineers?

Unfortunately, we failed to communicate this internally to our marketing team, and externally to reviewers at launch.

Indeed very conveniently unfortunate that you failed to do so consecutively for months.

some were disappointed that we didn't better describe the segmented nature of the architecture for that last 1GB of memory.

I would guess that more were disappointed that they unknowingly got a potentially crippled component...

Our only intention was to create the best GPU for you.

I totally believe you Mr. Jen! And I bet those shareholders want exactly the same...

This new feature of Maxwell should have been clearly detailed from the beginning.

Yay for new "features"!

We'll do a better job next time.

Oh God, no. What will be next? Shaders that only work with a monthly subscription? "Geforce Ultra Boost 9001! Get the extra power only when you really need it!"

 

[Bugsy004]

Best Mr. Jen & nvidia explained!!!

 

[Xzibit]

The miscommunication was between engineering and technical-marketing (the team that writes the Reviewer's Guide). Software is another department.

Yes, That's what is questionable. Engineering would have been done with their part long time ago and would have handed it off to software team long before the Reviewers Guide would be even conceived. Software team would have had months developing an appropriate driver and done memory allocation optimization that would differ from the 980 leading up to the release. One would think engineering and software teams would have input on the guide to what's possible or not, features and differences since the 980 and 970 were in the same guide. Just makes it seam like it was a copy and paste from the 980 when both engineering and software teams knew it was different from the start.

They are making it seam like everything is put together at last minute. We all know the hardware goes through months of testing before its even stable.

The break down would have been from all departments to the marketing team if that's plausible.

 

[xfia]

so that is the word! it is a great new feature they forgot to market! seems more like the new feature was for more efficient binning to raise that profit margin.

how does the letter even make since.. they say they figured out how to add more memory in a way like it was so hard and had to spend millions in research on it.

maybe they should talk to asus and ask them how they fit full 4gb on a 750ti if they are so clueless.

i dont know what to say to whoever believes this trash but I will say it will be interesting to see who stands up to challenge this architectures great new feature.

 

[john_]

Can you spot the sentence about false specs on ROPs and cache? No? I bet the lawyers they where also looking for those two words in the text.

 

[Ahhzz]

This isnt Blizzard.

OMG! someone recognized

 

[MustSeeMelons]

It wont happen again they say? Yesterday I was playing Dying Light (@ 1920x1200) and at one moment the card started using more than 3.5 Gb (Afterburner reported ~3900Mb) I got frame drops and stuttering, I got a feeling this is going to happen again and more often. My jump-ship finger is already tingling a bit.

 

[Fluffmeister]

What a disaster, and AMD still had to lower their prices.

 

[rtwjunkie]

"it wont happen again" "next time well do better" lol there is no next time nvidia, no way im gonna buy nvidia ever.

So you would deliberately limit yourself to a less capable card in the future (assuming correctly that at some point again Nvidia will have the top performing card again), even if it was cheaper? That doesn't make sense. I would think most logical people would try to buy the best performing card they can, no matter the brand.

Moving on: I have to say, it's obvious most people here are not familiar with how big companies work (rightly or wrongly is immaterial, it's reality). It's not uncommon at all for departments to not communicate thoroughly with each other, each in their own little turf, and who think and act as if they are THE central role of the company, with other departments merely supporting players. And no matter the product, the department that gets left out of the loop and looked down on the most? Marketing.

 

[EarthDog]

So you would deliberately limit yourself to a less capable card in the future (assuming correctly that at some point again Nvidia will have the top performing card again), even if it was cheaper? That doesn't make sense. I would think most logical people would try to buy the best performing card they can, no matter the brand.

Moving on: I have to say, it's obvious most people here are not familiar with how big companies work (rightly or wrongly is immaterial, it's reality). It's not uncommon at all for departments to not communicate thoroughly with each other, each in their own little turf, and who think and act as if they are THE central role of the company, with other departments merely supporting players. And no matter the product, the department that gets left out of the loop and looked down on the most? Marketing.

Quoted For Truth....

 

[Casecutter]

More like an explanation worded to sound like an apology

"More like a justification for a cover-up worded to sound like an apology"

This all might have been contrite... except for the fact the GM204 had "engineered" in the "Memory Crossbars" to connect and unused memory controller to still use the L2 that was orphaned. This just opened up the sink-hole, did no one in PR help this knucklehead.

 

[THE_EGG]

Buy a table and and 4 chairs,
Some one sits on the fourth chair,
It collapses.
"Well we were originally going to sell it with only three chairs, but we found an innovative new way to add a fourth. The last one is made of cardboard and bubble gum. We thought you'd be excited!"

Quite possibly the best analogy I've seen yet haha.

Though I thought I'd add to that and say that the fourth chair is still useable as the seat base still elevates one's butt off the ground however only at about 1/7th of the comfort of the rest of the chairs. Be happy that one's butt is not on the ground.

 

[RejZoR]

The miscommunication was between engineering and technical-marketing (the team that writes the Reviewer's Guide). Software is another department.

If they are unable to communicate something as basic as this, one starts to wonder what else they intentionally fucked up and no one at NVIDIA even knows about it, because there were more of these "communication errors" within their own company...

 

[Uplink10]

I don`t believe this guy, can`t he say "We gimp graphic cards to reduce their performance but this time we burned ourself because we gimped them too much". Do they really want to avoid class action lawsuit by saying we gimped 300$ graphic card less than usual but we still gimped it too much for all features to work normally?
Who makes this stuff up?

 

[EarthDog]

"We gimp graphic cards to reduce their performance but this time we burned ourself because we gimped them too much"

If that was what happened, he would have probably said that... but, that isn't what happened (gimped them "too" much).

 

[newtekie1]

So you´re not willing to buy amd because of doing the same thing nvidia has done... then that doesn´t make any sense, I think we should make notice us as a good and power consumers and just skip this generation to let them know we have the control and not them...

I never said I wouldn't buy AMD. I was just pointing out the hypocrisy of saying they won't buy nVidia again because of this situation because AMD has done worse and never even apologized.

When it comes to my purchases, I'll buy the best card for the money, I don't really care who it comes from.

Paging out to RAM will increase access times for non-critical stuff, but it also, to the best of my knowledge, does not create memory contention. If I understand correctly, a XOR situation occurs when one or the other segment needs to be written or read from, it causes a scenario where the other has to wait. This means any request to the 512MB segment has the possibility of harming the performance of the main 3.5GB segment.

EDIT: Either way, the GTX 970 will be a more driver dependent card, as Nvidia will probably have to spend more time working on how to manage this memory architecture with certain games.

The way I understand it is before anything can even be accessed by the GPU, it has to be brought back into VRAM from system RAM then accessed. This is a lot slower than the wait cycle to access the non-critical partition. Basically the main VRAM has to page something else out to system RAM(because it is already full), then read the data it wants into VRAM. That is a long process compared to the wait time to directly access the non-critical area.

No, they're disappointing because they unknowingly got a cut down part...

Actually we already knew it was a cut down part.

If they are unable to communicate something as basic as this, one starts to wonder what else they intentionally fucked up and no one at NVIDIA even knows about it, because there were more of these "communication errors" within their own company...

You obviously have no clue how hard it is to communicate technical shit to a marketing team. Almost everyone in this forum probably has more technical knowledge than anyone on a marketing team. If you tell the marketing team the card has 4GB, but 0.5GB of it is partitioned in a way that it is non-priority because the way the crossbar works now, you can disabled part of the L2, but still keep the memory controller associated with that L2 portion working by linking it over to the adjacent L2 portion, and because of this configuration the card has 64 ROPs but only 56 will be used because the other 8 ROPs are also accessed over this link to the adjacent L2 and it would actually hurt performance to use those extra 8 ROPs, even though they are enabled and technically could be used...

The marketing team is going to look at you with a blank stare and say "So the card is a 4GB card with 64 ROPs."

I don`t believe this guy, can`t he say "We gimp graphic cards to reduce their performance but this time we burned ourself because we gimped them too much".

Ironically, it was actually their attempt to not gimp the card as much as previous generations that bit them in the ass. If they did it the way they did in the past, the GTX970 would have been a 224-Bit 3.5GB card. However, the new method allowed them to not have to disabled that 32-bit memory controller and not loose the 0.5GB of memory that they normally would have had to.

 

[heydan83]

I never said I wouldn't buy AMD. I was just pointing out the hypocrisy of saying they won't buy nVidia again because of this situation because AMD has done worse and never even apologized.

When it comes to my purchases, I'll buy the best card for the money, I don't really care who it comes from.



The way I understand it is before anything can even be accessed by the GPU, it has to be brought back into VRAM from system RAM then accessed. This is a lot slower than the wait cycle to access the non-critical partition. Basically the main VRAM has to page something else out to system RAM(because it is already full), then read the data it wants into VRAM. That is a long process compared to the wait time to directly access the non-critical area.



Actually we already knew it was a cut down part.



You obviously have no clue how hard it is to communicate technical shit to a marketing team. Almost everyone in this forum probably has more technical knowledge than anyone on a marketing team. If you tell the marketing team the card has 4GB, but 0.5GB of it is partitioned in a way that it is non-priority because the way the crossbar works now, you can disabled part of the L2, but still keep the memory controller associated with that L2 portion working by linking it over to the adjacent L2 portion, and because of this configuration the card has 64 ROPs but only 56 will be used because the other 8 ROPs are also accessed over this link to the adjacent L2 and it would actually hurt performance to use those extra 8 ROPs, even though they are enabled and technically could be used...

The marketing team is going to look at you with a blank stare and say "So the card is a 4GB card with 64 ROPs."

So if they would have been explained well, every 970 box would have a legend that reads "3.5 GB + 500!! MB that works weird" who would have purchased something like that?, it would be a 780 with 500 MB extra and no one would buy it.

And also the basis of any company is the communication between every of the work areas, what a coincidence that this communication problem happened when there was something bad with the card, why they don´t communicate the 980 was 3GB by mistake and then everyone realize it was actually a 4GB card.

Ok yes it is difficult to communicate somthing like this to the marketing team.... so what they did? maybe they lie the marketing team and just tell them the card is a 4GB card....