Editorial NVIDIA's 20-series Could be Segregated via Lack of RTX Capabilities in Lower-tier Cards

[D1RTYD1Z619]

maybe a 2080 RTX at 1500$ and then a 2080 GTX at 750$ [with 20$ mail in rebate]

It would probably just be 10 bucks of Fortnite credit.

 

[hat]

As I understand it, process defects at the foundry are largely responsible for product segmentation. For instance (pulling numbers out of my ass here, because I don't know the chip names that well) they may be "trying" to make GP100 chips at the foundry, and the perfect chips get put on Quadros and such, and the "defective" chips (which they may call GP102 or something) may get slapped on a Titan or high end GTXwhatever, with some shader units disabled or something, wherever the defect was. This theory seems to be supported by that one Quadro in the current lineup that is faster than even the current Titan in games, by a margin that isn't insignificant.

 

[TheoneandonlyMrK]

As I understand it, process defects at the foundry are largely responsible for product segmentation. For instance (pulling numbers out of my ass here, because I don't know the chip names that well) they may be "trying" to make GP100 chips at the foundry, and the perfect chips get put on Quadros and such, and the "defective" chips (which they may call GP102 or something) may get slapped on a Titan or high end GTXwhatever, with some shader units disabled or something, wherever the defect was. This theory seems to be supported by that one Quadro in the current lineup that is faster than even the current Titan in games, by a margin that isn't insignificant.

That's not just a theory , it's how the majority of chips are made, that and with spare overprovisioned part's to step up when another parts faulty(mostly Amds tactic) in chip.
Without such tactics 20-80% of every wafer could be wasted and the cost of the working ones would be astronomical.

 

[StrayKAT]

Glad to see you figured it out too @Raevenlord

I knew this when I saw Jensen shout TEN GIGA RAYS like a fool. He tried to pull a Steve Jobs on people. Everything was amazing, fantastic, never seen before, first time you could get proper use out of this new tech... it was like watching an Apple keynote.

Except with Nvidia, what they showed were jerky tech demoes at 30 FPS and a huge blurry mess of a dude dancing at the end. Oh yeah, it had reflections, too.

Its hilarious to see people on TPU echoing that RTRT is the next best thing. As if Nvidia re-invented the wheel. I guess this generation separates the fools from the realists.

Except instead of like Steve Jobs, he replaced the turtleneck with a leather jacket. He probably thinks he's Tony Stark.

 

[Captain_Tom]

This.
It looks like TU102/4 was specifically made for Quadro cards and making a RT and Tensor core-less GPU just for gaming cards wasnt finacially viable if 7nm is just around the corner.

Exactly. They are launching both the 2080 and the 2080 Ti at the same time, and neither are the full die - this is a 10 month series at most haha.

 

[TheGuruStud]

Except instead of like Steve Jobs, he replaced the turtleneck with a leather jacket. He probably thinks he's Tony Stark.

Hahahaha, I forgot just how big of a douche Jobs was. It all makes sense, now. It's like requisite apparel to milk your dummy customers.

 

[TheDeeGee]

So 2060 should sell for $100 then?

 

[Manoa]

- GP100’s SM incorporates 64 single-precision (FP32) CUDA Cores. In contrast, the Maxwell and Kepler SMs had 128 and 192 FP32 CUDA Cores, respectively.
that ! also in linux on nouveau you get full speed FP64 without artificial crippling from the driver on the 780 Ti. what a card that !, it can run on all windows 5 and nouveau. now I know why my 780 Ti can protect 1 TB of data in PAR2 in 1 hour !

 

[Krzych]

Do those new RTX card support 10bit colour or still 8bit like rest of Nvidia Geforce series?

Unlikely. There is a pretty good explanation for that on Nvidia website:

"NVIDIA Geforce graphics cards have offered 10-bit per color out to a full screen Direct X surface since the Geforce 200 series GPUs. Due to the way most applications use traditional Windows API functions to create the application UI and viewport display, this method is not used for professional applications such as Adobe Premiere Pro and Adobe Photoshop. These programs use OpenGL 10-bit per color buffers which require an NVIDIA Quadro GPU with DisplayPort connector."

I don't think there is any reason for this to change, they have no interest in bringing pro features into mainstream.

 

[TheGuruStud]

Unlikely. There is a pretty good explanation for that on Nvidia website:

"NVIDIA Geforce graphics cards have offered 10-bit per color out to a full screen Direct X surface since the Geforce 200 series GPUs. Due to the way most applications use traditional Windows API functions to create the application UI and viewport display, this method is not used for professional applications such as Adobe Premiere Pro and Adobe Photoshop. These programs use OpenGL 10-bit per color buffers which require an NVIDIA Quadro GPU with DisplayPort connector."

I don't think there is any reason for this to change, they have no interest in bringing pro features into mainstream.

I don't think this is nvidia or AMD, but rather adobe and their delusions of grandeur, unless it's collusion for segmentation (which sounds illegal...not that anyone has given two shits in 20 yrs about megacorp legality).

 

[TheoneandonlyMrK]

Unlikely. There is a pretty good explanation for that on Nvidia website:

"NVIDIA Geforce graphics cards have offered 10-bit per color out to a full screen Direct X surface since the Geforce 200 series GPUs. Due to the way most applications use traditional Windows API functions to create the application UI and viewport display, this method is not used for professional applications such as Adobe Premiere Pro and Adobe Photoshop. These programs use OpenGL 10-bit per color buffers which require an NVIDIA Quadro GPU with DisplayPort connector."

I don't think there is any reason for this to change, they have no interest in bringing pro features into mainstream.

Pro features , i hardly call 10bit pro it comes in handy for HDR and every amd card out now can do it easily with a compatible monitor over hdmi or displayport.
Also such segregation offends my sensibility, yeah some pro features can be kept for quadro but that's not one i agree with.

 

[notb]

I remember the days when you used to be able to trick the driver to think a GeForce card was a Quadro card by faking a different device id.

Why do that?

I wouldn't be surprised if ECC is supported but, they just put non-ECC DRAM on it for the consumer.

Of course the chip itself is compatible with ECC. But why would you want gamers to pay for ECC RAM?

2x ? 2x what ? Price ?

2x performance in best case scenario (tensor cores taking over AA). There are a few games which already support this.
I seems the average improvement is around 50% (not bad, right?)

- GP100’s SM incorporates 64 single-precision (FP32) CUDA Cores. In contrast, the Maxwell and Kepler SMs had 128 and 192 FP32 CUDA Cores, respectively.

GP100 has 64 FP32 and 32 FP64 cores. It's the same with GV100.
Also, why the drama if you don't care that much? You've missed today's medicine or what?

 

[Manoa]

I think the idea of dedicated anti aliasing tensors is very good, don't need to bother with special anti aliasing parts in the rasterization process
I have nothing against ECC for gamers, I think it also a good idea, why should gamers have shit hardware ?
when you "quadrify" a geforce card you unlock 2 things: bypass artificial FP64 limitation of nvidia driver (which limits to 1/4 FP64 operations of the hardware ability), enable hardware virtualization PCI passthrough and bypass code 43 of nvidia driver
192 FP32 vs 64 FP32 units ? is that realy something that people shouldn't care about ?
if nvidia didn't limit FP64 capacity of geforce cards, and the speed was there, developers would be using more of it, high precision computations translates to higher quality graphics, it simple as that, wouldn't even need RTRT to look good

 

[notb]

I think the idea of dedicated anti aliasing tensors is very good, don't need to bother with special anti aliasing parts in the rasterization process

These are not "dedicated anti aliasing tensors". Tensor cores simply are good at a certain type of operations - just like the general "CUDA" cores.

I have nothing against ECC for gamers, I think it also a good idea, why should gamers have shit hardware ?

Why do you think ECC is so important?

192 FP32 vs 64 FP32 units ? is that realy something that people shouldn't care about ?

Exactly, they shouldn't. It's a product. Nvidia tells you what it's good for. Then there are reviews which check if those claims are true.
If the products does what you want, buy it. If not, buy something else.
You're simply overconcerned about some magic numbers on the spec sheet.

developers would be using more of it, high precision computations translates to higher quality graphics, it simple as that, wouldn't even need RTRT to look good

Actually, no. High precision doesn't translate into higher quality graphics. For graphics you need many low-precision operations. This is exactly the reason why GPUs exist.

 

[sweet]

Unlikely. There is a pretty good explanation for that on Nvidia website:

"NVIDIA Geforce graphics cards have offered 10-bit per color out to a full screen Direct X surface since the Geforce 200 series GPUs. Due to the way most applications use traditional Windows API functions to create the application UI and viewport display, this method is not used for professional applications such as Adobe Premiere Pro and Adobe Photoshop. These programs use OpenGL 10-bit per color buffers which require an NVIDIA Quadro GPU with DisplayPort connector."

I don't think there is any reason for this to change, they have no interest in bringing pro features into mainstream.

Quick reminder that AMD has these features in their Vega FE though.

 

[TheoneandonlyMrK]

Quick reminder that AMD has these features in their Vega FE though.

Polaris and vega both can output to upto 12bit afaik and have tried

 

[qubit]

@Raevenlord I'd love to see a tiddly RT 1030 type card with full ray tracing capabilities present and correct, performing at 1fps or less, just for the fun of it. And yes, I'd be that nerd that bought it just to watch that slide show. After all, I did that with the GT 1030 to complement my GTX 1080, but was actually surprised that it performs as well as it does for such a little GPU.

 

[hat]

GTX580 seems twice as fast... but it beats the much older GTX285. Still, I was rocking a GTX260 at one time... admirable that such low end "garbage" is actually able to perform somewhat.

 

[sweet]

Polaris and vega both can output to upto 12bit afaik and have tried

I was talking about 10bit in OpenGL Adobe products. Vega frontier is the cheapest solution for that at this moment.

 

[ghazi]

As I understand it, process defects at the foundry are largely responsible for product segmentation. For instance (pulling numbers out of my ass here, because I don't know the chip names that well) they may be "trying" to make GP100 chips at the foundry, and the perfect chips get put on Quadros and such, and the "defective" chips (which they may call GP102 or something) may get slapped on a Titan or high end GTXwhatever, with some shader units disabled or something, wherever the defect was. This theory seems to be supported by that one Quadro in the current lineup that is faster than even the current Titan in games, by a margin that isn't insignificant.

While what you're describing definitely is a practice (a la x80 vs x70, or i7 vs i5 etc) GP100 and GP102 are a different matter; NVIDIA made two different high-end chips due to the additional die size FP64 and other niche professional features require. Additionally GP100 uses HBM2 while GP102 uses GDDR5X. GP102 is significantly cheaper for NVIDIA to produce.

 

[AsRock]

Why does this article act like this is new info? Everyone knows the 2060 is very unlikely to have RT cores - Nvidia wants its sheep to pay $600 cards from now on.

Furthermore the x60 series is now basically made for laptops first, and the new RT cores are HORRIBLY inefficient. It makes substantially more sense to create a Turing die that lacks RT and Tensor cores so they can make it a tiny <200mm^2 die that only uses 50-100w.

Yes but they are marketing the wholle range as RTX when really they are not. As they know full well the low end ones will sell as company's like best buy and such will sell them with a system telling them they have ray tracing, which by time it becomes a thing it be way to late for the people who got the cards to do any thing about it.

Another nVidia ripp off.

 

[notb]

Polaris and vega both can output to upto 12bit afaik and have tried

10-bit in games - not in Windows API. It's exactly the same with Nvidia cards.

 

[efikkan]

"NVIDIA Geforce graphics cards have offered 10-bit per color out to a full screen Direct X surface since the Geforce 200 series GPUs. Due to the way most applications use traditional Windows API functions to create the application UI and viewport display, this method is not used for professional applications such as Adobe Premiere Pro and Adobe Photoshop. These programs use OpenGL 10-bit per color buffers which require an NVIDIA Quadro GPU with DisplayPort connector."

I don't think there is any reason for this to change, they have no interest in bringing pro features into mainstream.

Their Linux drivers does support 10-bit per channel for all recent GeForce cards though.

 

[Aquinus]

Why do that?

Because it actually unlocked some of the professional features that weren't actually a hardware limitation at the time because a lot of things were blocked by the driver. Remember ever unlocking pixel or vertex pipelines via software?

Of course the chip itself is compatible with ECC. But why would you want gamers to pay for ECC RAM?

...but why does it support ECC even if it's not being used for the consumer? Usually that's because it has an application where ECC is required. Hence me jumping to the conclusion that there are just professonal cards that were slightly altered for the consumer because they had already been built. It also could be the case that the demand for these cards wasn't as high as they had anticipated and are repurposing these chips instead.

The big thing is if these are professional cards with a lot more hardware in them, how is that going to impact clock speed and heat? Usually when this kind of thing happens, things don't live up to expectation and that's usually because despite all of the compute hardware, it's not really giving the GPU an advantage when it comes to rendering stuff in games.

Their Linux drivers does support 10-bit per channel for all recent GeForce cards though.

Too bad that you're practically selling your soul by using nVidia's closed source drivers. I'm still waiting for them to release their firmwares so nouveau can suck a little less but, obvious nVidia has no intent to make the open source community happy.

 

[notb]

Remember ever unlocking pixel or vertex pipelines via software?

No...?
It's been 10 years since I stopped overclocking, modding, unlocking and so on. It was fun when I was a teenager, but really seem to have been a waste of time. I should have just spent more time outside or learn German instead.

...but why does it support ECC even if it's not being used for the consumer? Usually that's because it has an application where ECC is required. Hence me jumping to the conclusion that there are just professonal cards that were slightly altered for the consumer because they had already been built. It also could be the case that the demand for these cards wasn't as high as they had anticipated and are repurposing these chips instead.

No. "Gaming" cards are not altered "pro" cards. It's just a chip which can do multiple things. Firmware and drivers tell him what to do.
The reason why gaming chips have ECC functionality included (but blocked) is fairly simple: there's nothing special about ECC. Pretty much all CPUs and GPUs sold today are ECC-compliant.
Manufacturers are using ECC as a factor in lineup segmentation, which really makes sense.
It's not like ECC is crucial for what most people do with their PCs and ECC RAM is more expensive.

Sure, we could suddenly make all computers use ECC.
But sooner or later someone would notice that consumers are paying a premium for a feature they don't need, so it must be a conspiracy by chip makers! Just look how people reacted to RTX.
And the overclocking crowd would soon decide that they miss the non-ECC times, because they could OC higher.