NVIDIA Maxwell Speculation Thread

[fbomber]

Using your logic would not the end user then have to pay Nvidia going forward when new drivers improve the "price/performance".

No. Completely different situations. False advertising is illegal. It´s even a crime in many countries. You can not lie to a consumer to make he/she buy your product, as it would be fraud.

Even if it was just a mistake by Nvidia, those that were mislead by the company´s statements which misrepresented the product and feel cheated should be compensated. Consumers have the right to know what they are buying.

 

[Dave Baumann]

Considering the amount of GPU brainpower on this thread it is amusing that people who actually use the cards spotted it rather than you guys.

Perhaps they actually use the cards rather than just talk forever on what is under the hood?

Rather a sad conclusion on how Beyond3D has gone downhill since Dave left I think.

Perfectly understandable that this would be "discovered" by end users rather than theoretical noodling around. For one, the tools for the end user have got better (or at least, more accessible) to be able to discover this type of thing. Second, the fundamental interconnects within a GPU are not the parts that are ever discussed, because largely they aren't necessary to know about; additionally ASIC "harvesting" characteristics often change between relatively minor architectural changes, or even from chip configurations within the same family.

 

[Rys]

They are making the consumer whole again if that's what they want.

As to having a right to know what you're buying, nobody is obliged to disclose the architectural details of their processor up front. You're within your rights to buy it and return it if it doesn't do what you were expecting for the money you paid, but having the details NV are in trouble over, as part of that expectation, is luxury.

The details are disclosed to reviewers because the enthusiast segment "sells" boards to other segments via recommendation. As soon as that stops being profitable for the vendor, disclosure will go away. This furore puts that in danger, in my opinion anyway.

 

[firstminion]

Using your logic would not the end user then have to pay Nvidia going forward when new drivers improve the "price/performance".

That's preposterous.

 

[firstminion]

They are making the consumer whole again if that's what they want.

As to having a right to know what you're buying, nobody is obliged to disclose the architectural details of their processor up front. You're within your rights to buy it and return it if it doesn't do what you were expecting for the money you paid, but having the details NV are in trouble over, as part of that expectation, is luxury.

The details are disclosed to reviewers because the enthusiast segment "sells" boards to other segments via recommendation. As soon as that stops being profitable for the vendor, disclosure will go away. This furore puts that in danger, in my opinion anyway.

I would say that the asymmetry puts the (theoretical) Memory Bandwidth value in jeopardy and that is a memory specification. http://www.geforce.com/hardware/desktop-gpus/geforce-gtx-970/specifications

We don't need to pander nvidia, if they don't want to disclose architectural details, don't disclose it. But if they do, bite the bullet.

 

[3dilettante]

I would say that the asymmetry puts the (theoretical) Memory Bandwidth value in jeopardy and that is a memory specification. http://www.geforce.com/hardware/desktop-gpus/geforce-gtx-970/specifications

The specifications can be argued as being derived from the physical capability of the DRAM chips and physical interface of the GPU, which are straightforward and physically true.
It's also unclear if the bandwidth number cannot be reached in certain situations. The articles on the subject focus on read bandwidth, but what if one partition has reads and the other has writes?

 

[Ryan Smith]

The articles on the subject focus on read bandwidth, but what if one partition has reads and the other has writes?

Then you can achieve the 224GB/sec. Though pulling that off in practice is very hard, which is why NV doesn't even activate the last segment until you need the extra VRAM.

(The write bus is shared, just like the read bus. So to get 224GB/sec you need to be able to co-issue reads and writes every last cycle. Otherwise if there's a block, you quickly lose bandwidth)

 

[pjbliverpool]

I remember when Microsoft promoted the Xbox 360 as a 1TFLOP console. I wonder what would have happened if people had been equally offended by that.

 

[dbz]

No. Completely different situations. False advertising is illegal. It´s even a crime in many countries. You can not lie to a consumer to make he/she buy your product, as it would be fraud.
Even if it was just a mistake by Nvidia, those that were mislead by the company´s statements which misrepresented the product and feel cheated should be compensated. Consumers have the right to know what they are buying.

Just out of interest, how often do these kind of things go to court? I seem to remember a similar situation (re: advertised features) regarding the 2900 XT's UVD status. Checked through a couple of old threads here ( and here), but the issue seemed to have just hand-waved away or magically disappeared

 

[snc]

Reading from last 512mb vram in gtx970 is 22GB/s, and pci express 3 is 16GB/s so no big difference, but to read from system ram we have to copy it to vram(16GB/s) and then read from vram, thats the difference ?

 

[Rys]

Right, bandwidth arguably doesn't even matter at that point. There are very few situations where, if the GPU needs something that isn't in its directly addressable memory, latency isn't what actually murders performance. It's nuanced (the two are inextricably linked), but finding things to do for thousands of clocks is very hard and will almost guarantee to starve the chip.

 

[kukreknecmi]

So for a 980, if Nvidia GPU wants to write memory(say uncached write), then it does this in 1KB granulity right? For a 32KB write , does this via (1KB x 8 channel) x 4 times , am i getting it right?

 

[LordEC911]

Nvidia has said they will give refunds to anyone who want them I think.

No. Nvidia never said that. A Director of Customer Care at Nvidia, asked GTX970 owners that want to return the card to contact the retailer and AIB, then if you are denied he will contact them on your behalf. There is no guarantee of anything being done for you.

 

[firstminion]

The specifications can be argued as being derived from the physical capability of the DRAM chips and physical interface of the GPU, which are straightforward and physically true.
It's also unclear if the bandwidth number cannot be reached in certain situations. The articles on the subject focus on read bandwidth, but what if one partition has reads and the other has writes?

"Physically true", but in jeopardy nonetheless. Ryan remembered that Nvidia doesn't even activate the last segment until the last moment. It feels a little crooked.

 

[Malo]

I don't know why I'd give up my 970. The performance and cooling is fantastic for the price.

 

[Esrever]

I don't know why I'd give up my 970. The performance and cooling is fantastic for the price.

It all depends on your perception on the basis of morality and such. Also hinges on if you view nvidia basically hiding a serious flaw directly effect you in anyways. To some people, it show a lack of respect for their own customers when a company lies like this. I recommended the 970 to a friend and even on his behalf I feel kind of bad even though the issue will likely never come up.

 

[Ryan Smith]

"Physically true", but in jeopardy nonetheless. Ryan remembered that Nvidia doesn't even activate the last segment until the last moment. It feels a little crooked.

Remembered? That information came from NVIDIA (Jonah) back on Sunday.

 

[Putas]

For many that don´t upgrade as often, a card that is not future-proof is not an option. And they eventually woulnd´t have bought the GTX 970 if they knew it has less ROPs or useable memory than what Nvidia announced.

I have doubts about this changing future-proofness of 970. Assuming 980 is balanced, 970 in general use does not have much need for more ROPs, bandwidth and capacity.

very clear case of false advertisement, they were very clear when said the card had 64ROPs and 256bit 4GB, while the 64 ROPs is 100% false, the rest also seems false to me, the correct is up to 224 bits memory for 3.5GB and UP to 32bit for ther other 512MB, and never 256bits speed (adding the 512MB to the 3.5GB drops the bandwidth, it doesn't add, like it would on a 980, so 256bits cannot be achieved)

The bits are actual connections, there is nothing like "up to 32 bits". You can argue it is 224 and proportionally lower bandwidth because the last MC cannot be used at once. The problem with through-output of last chunk is (if I understand it ) the complete lack of L2 cache - it has nothing to do with bits.

Wasn't the rumor that R380 would couple HBM and GDDR5 in unison...?

Jeez, are they gonna reuse codenames like that?

 

[Blazkowicz]

Just out of interest, how often do these kind of things go to court? I seem to remember a similar situation (re: advertised features) regarding the 2900 XT's UVD status. Checked through a couple of old threads here ( and here), but the issue seemed to have just hand-waved away or magically disappeared

I don't remember that.
What about the video engines on anything pre geforce 8600GT/8500GT/8400GS : there was a video engine advertised but it was pretty useless, such as MPEG2 assistance in specific video players and partial assistance in H264 and VC-1 which maybe hardly anyone ever used. It took years for video decoders to get support in video players, flash plugin etc. (DXVA and VDPAU) by which time the early video engines were 100% useless and deprecated.

 

[HMBR]

The bits are actual connections, there is nothing like "up to 32 bits". You can argue it is 224 and proportionally lower bandwidth because the last MC cannot be used at once. The problem with through-output of last chunk is (if I understand it ) the complete lack of L2 cache - it has nothing to do with bits.

32bits is the maximum for the last 512MB, because of the way they have disabled the ROP/L2 (which normally would make them also disable that MC) to which it should be connected and it operates separately from the other MCs, so the maximum bandwidth is just for one MC (32bits) while reading/writing to that, while the other 3.5GB can use the full 224 bits from the 7 MCs combined (and the 980 can use the full 8 MCs = 256bits)

they also advertised the card as having 64 ROPs and 2MB of l2, maybe they didn't write it on the box or nvidia.com, but they sent these specs for sites like Anandtech, TomsHardware and so on, which they use to advertise their products, and they fully expected these sites to publish these false specs... so even if you think the 224GB/s is debatable, the ROPs/L2 are clear;

as for them helping with refunds, it looks to me like they are helping forum members/social media active persons to get their refunds, seems fine, but maybe something more official would be better for the others.