Poll: Tensor (TPU) Cores on consumer GPUs: Boon or Bust?

Tensor Cores, Boon or Bust?

  • Bust - it has failed so far, and doesn't appear to have a bright future.

    Votes: 22 31.9%
  • Boon - Already providing notable benefits.

    Votes: 5 7.2%
  • Future Boon - questionable today, but expect future benefit.

    Votes: 42 60.9%

  • Total voters
    69
  • Poll closed .

Snowdog

[H]F Junkie
Joined
Apr 22, 2006
Messages
11,262
Right now machine learning is all the rage. TPU, Tensor, ML cores are being added to GPUs and even Mobile SoCs.

For GPUs NVidia has launched RTX cards with RT cores and Tensor cores. The latter used to "aid" Ray Tracing by running Denoising networks and DLSS networks.

But results are mixed. BFV devs said they don't used Tensor cores to denoise. It also seems like Q2-RTX uses some kind of Temporal algorithm to denoise, not Tensor cores. DLSS has failed to deliver any benefit more than a simple resize and process, but requires specific training.

So how do you see Tensor Cores on Consumer GPUs. A benefit, a bust, or potential future benefit?
 
Bust or Future Boon...I could see it going either way (leaning toward bust), but I don't see there being any notable benefits on the consumer side of things right now.
 
I think it's definitely a future boon. Just like most new tech.
 
Future Boon. I think it is still too soon to call it bust. I believe there is not enough of a notable benefit currently for me to justify a hardware upgrade.
 
DXR is cool tech but still a long way to go for mainstream. One expensive card can run it at acceptable frames at 4k in one game and it's partial RT, not even full scene.
DLSS is an utter joke.

Future boon if they keep pouring money in it. At this point it's far from ready.
 
I am leaning towards bust, mostly because Tensor cores are highly specialized and judging from the dev feedback nobody quite knows how to integrate/what to make of them.
That could change in the future I suppose -- but I doubt it. I am mostly basing this on those specialized PhysX cores which just disappeared as well whereas PhysX itself stuck around.
 
I am leaning towards bust, mostly because Tensor cores are highly specialized and judging from the dev feedback nobody quite knows how to integrate/what to make of them.
That could change in the future I suppose -- but I doubt it. I am mostly basing this on those specialized PhysX cores which just disappeared as well whereas PhysX itself stuck around.

Leaning that way as well. They even seem unnecessary for Ray Tracing. Since there are alternatives for noise reduction, and DLSS failed completely to deliver.

They may interest developers some day that find use for Neural networks in games, but so far they look like a solution in search of a problem.
 
Leaning that way as well. They even seem unnecessary for Ray Tracing. Since there are alternatives for noise reduction, and DLSS failed completely to deliver.

They may interest developers some day that find use for Neural networks in games, but so far they look like a solution in search of a problem.

I am also wondering if Tensor cores are just cheap/simple enough to plop down on a GPU(chip) that Nvidia put them there to see if anything sticks.

Judging by Google's 15-month concept-to-deployment TPU (https://cloud.google.com/blog/produ...k-at-googles-first-tensor-processing-unit-tpu) it's not entirely bullshit to say Tensor cores are almost trivial compared to other HW development. Definitely easier than trying to innovate anywhere else.

I work in CPU/chip HW design and verif -- 15 months is ridiculously short.
 
I am also wondering if Tensor cores are just cheap/simple enough to plop down on a GPU(chip) that Nvidia put them there to see if anything sticks.

Judging by Google's 15-month concept-to-deployment TPU (https://cloud.google.com/blog/produ...k-at-googles-first-tensor-processing-unit-tpu) it's not entirely bullshit to say Tensor cores are almost trivial compared to other HW development. Definitely easier than trying to innovate anywhere else.

I work in CPU/chip HW design and verif -- 15 months is ridiculously short.


Tensors could be there to share dies with Pro usage. Big neural nets run on a bank of GPUs is a real usage for Tensor cores.

Little Neural Nets at home, seems much less required.
 
Tensors could be there to share dies with Pro usage. Big neural nets run on a bank of GPUs is a real usage for Tensor cores.

Little Neural Nets at home, seems much less required.

Exactly they are there because unless something changes NVs highest performing margin sectors are AI and automotive AI. Automotive does require different silicon... but the data centre stuff does not.

Volta had tensor cause it sells AI data centre cards. Turing is just updated volta. It was not designed for gaming first... some choices may have been for gaming of course. But the entire design was not consumer bent.

As Snowdog says it seems like a solution in search of a problem. I still believe based on how the "RT cores" work they are in fact using the tensor hardware. However even if we believe NV and they designed a RT core that isn't related to Tensor in any real way. Again the game devs don't seem to be using tensors for denoise as NV suggest.

I'm not sure tensors have much of a future for consumers... they will be on NVs silicon however for as long as they can keep selling them to data centres. If some better custom solution from Google directly or Intel comes a long I don't suspect they are going to abandon the Shader Core+Tensor core designs.
 
Tensors could be there to share dies with Pro usage. Big neural nets run on a bank of GPUs is a real usage for Tensor cores.

Little Neural Nets at home, seems much less required.

The big servers are there for training the neural net. Need much less horsepower for inferencing on a trained network.

Maybe one day game AI will run on tensors instead of hard coded rules. Short term I vote bust though. I think it’ll be a while before we come up with killer gaming applications for ML.

Have to give nVidia props for DLSS though. At least they came up with “something” to get the ball rolling.
 
The big servers are there for training the neural net. Need much less horsepower for inferencing on a trained network.

I get that. But what are the applications for the home user of those trained network that are actually beneficial, and can't already be done (possibly better) using CPU or GP-GPU code.

Have to give nVidia props for DLSS though. At least they came up with “something” to get the ball rolling.

I'd give them props if it actually delivered. After the initial reveal I thought this would be bigger than Ray Tracing initially. This looked like it had so much promise until the quality was shown to be very questionable. Essentially it's no better that traditional resolution scaling, and the touted DLSS 2x, that was supposed to approach 64x Super Sampling AA, never actually showed up during the months following release.
 
I think it was secondarily just an excuse to get Tensors out on consumer hardware and kick start some use. DLSS isn't impressive to me, and I don't really see anything else that has used them to date. But maybe someone will think of something, since most people don't run industrial AI networks at home.

I think the primary reason was to help justify higher costs, since 2000 series isn't really all that much faster at rasterization than the 1000 series.
 
Forty percent isn't a lot?

I have no idea where you get this number but I won’t be the one to burst your bubble.

Golden is right, it’s about 40% if it’s not CPU limited. A lot of reviews early on started with a 1080 (or lesser) around 60fps and by the time they got to the 2080ti it was cpu limited. The thing is so fast it can be easy to do.

Personally I didn’t factor RTX into my choice to buy a 2080ti. I wanted the speed for VR. I was excited for DLSS2X though which never actually materialized...
 
Last edited:
I have no idea where you get this number but I won’t be the one to burst your bubble.

lol 2080ti it's at least 40% faster than 1080Ti just check every review out there, and even more when comparing overclock vs overclock... 2080Ti it's another level of performance.. that shit can even be CPU limited on most games at 1440P@120+FPS ultra settings..
 
1080ti vs 2080ti ... 40% faster even if we all agree those numbers are 100% correct. Is that even a fair comparison ?

The 2080ti launched at a price 30-40% higher then the 1080ti launch msrp.

The 2080 launched at the same MSRP as the 1080ti... and the performance there is basically identical not counting RTX stuff in the 3 or 4 games and couple demos you can turn it on in.

Bottom line price to performance (not counting RT stuffs) nothing changed from one generation to the next. Gamers spending $700 last generation got the exact same performance spending $700 this generation. The only value add is a bunch of half baked features that may or may not have a future. The 2080ti is a great card no doubt... but its founders edition MSRP was identical to last generations Titan card, and its not 40% faster then that card. Yes its faster by 10-25% depending on the test don't get me wrong its a nice bump up... but lets compare Apple$ to Apple$. 1080ti should really be compared to the 2080... and that is neck and neck. 2080ti is faster then the titan at the same MSRP for sure... but its not 40%.

NV slide their product stack up. The main issue a lot of people have with the 1080ti -> 2080ti comparisons are not based on performance but cost. The 2080ti is $300-500 more expensive depending if you where in day one on a founders edition. So ya it better darn well be 40% faster as its basically 40% more expensive. lol
 
I actually like DLSS - when used with high enough resolution. I prefer how Shadow of the Tomb Raider looks with DLSS 4K over native 4K and appreciate the performance boost. With SoTR DLSS at 4K gives it a very stable image that I feel is still sharp but not aliased. No shimmering issues. At any lower res it just looks blurrier. Nvidia themselves say that the higher the res the better DLSS works as it has more pixels to work with. That said I want to see more options as 1080p on 4K resolution should also be usable with 1:1 integer scaling instead of the blurry mess AMD and Nvidia offer on PC.

I don't know enough about Tensor core use cases to say whether they are good to have or not. Maybe they just haven't found the killer use case for it when it comes to games. Maybe we could have them handle AI processing or something in the future, who knows.

I am sold on raytracing tech though. Next few generations will make it a norm and hopefully with the performance increases needed too. As a 2080 Ti owner I am still excited about being able to use it right now even if it incurs a hefty performance hit. I don't mind too much as I've never been one to chase the absolute highest frame rates and I don't play competitive multiplayer shooters.
 
1080ti vs 2080ti ... 40% faster even if we all agree those numbers are 100% correct. Is that even a fair comparison ?

The 2080ti launched at a price 30-40% higher then the 1080ti launch msrp.

The 2080 launched at the same MSRP as the 1080ti... and the performance there is basically identical not counting RTX stuff in the 3 or 4 games and couple demos you can turn it on in.

Bottom line price to performance (not counting RT stuffs) nothing changed from one generation to the next. Gamers spending $700 last generation got the exact same performance spending $700 this generation. The only value add is a bunch of half baked features that may or may not have a future. The 2080ti is a great card no doubt... but its founders edition MSRP was identical to last generations Titan card, and its not 40% faster then that card. Yes its faster by 10-25% depending on the test don't get me wrong its a nice bump up... but lets compare Apple$ to Apple$. 1080ti should really be compared to the 2080... and that is neck and neck. 2080ti is faster then the titan at the same MSRP for sure... but its not 40%.

NV slide their product stack up. The main issue a lot of people have with the 1080ti -> 2080ti comparisons are not based on performance but cost. The 2080ti is $300-500 more expensive depending if you where in day one on a founders edition. So ya it better darn well be 40% faster as its basically 40% more expensive. lol

Not only that, you used to get that extra 35-40% for "free" just by upgrading generation to generation (see 980Ti to 1080Ti, etc.). Now Nvidia expects you to pay an extra 40% for your extra 40% performance with the half-baked features, etc.
 
1080ti vs 2080ti ... 40% faster even if we all agree those numbers are 100% correct. Is that even a fair comparison ?

The 2080ti launched at a price 30-40% higher then the 1080ti launch msrp.

The 2080 launched at the same MSRP as the 1080ti... and the performance there is basically identical not counting RTX stuff in the 3 or 4 games and couple demos you can turn it on in.

Bottom line price to performance (not counting RT stuffs) nothing changed from one generation to the next. Gamers spending $700 last generation got the exact same performance spending $700 this generation. The only value add is a bunch of half baked features that may or may not have a future. The 2080ti is a great card no doubt... but its founders edition MSRP was identical to last generations Titan card, and its not 40% faster then that card. Yes its faster by 10-25% depending on the test don't get me wrong its a nice bump up... but lets compare Apple$ to Apple$. 1080ti should really be compared to the 2080... and that is neck and neck. 2080ti is faster then the titan at the same MSRP for sure... but its not 40%.

NV slide their product stack up. The main issue a lot of people have with the 1080ti -> 2080ti comparisons are not based on performance but cost. The 2080ti is $300-500 more expensive depending if you where in day one on a founders edition. So ya it better darn well be 40% faster as its basically 40% more expensive. lol

Not only that, you used to get that extra 35-40% for "free" just by upgrading generation to generation (see 980Ti to 1080Ti, etc.). Now Nvidia expects you to pay an extra 40% for your extra 40% performance with the half-baked features, etc.

Still raining where you guys are huh?
 
Still raining where you guys are huh?

You can take it on the chin and pay the extra money if you'd like. They'll pry it out of my cold dead hand (at least until I find an actual use for the "features" which there isn't at this point).
 
Still raining where you guys are huh?

It's not rain, Just the lingering tears over the slow grinding end of "Moore's Law".

Not only that, you used to get that extra 35-40% for "free" just by upgrading generation to generation (see 980Ti to 1080Ti, etc.). Now Nvidia expects you to pay an extra 40% for your extra 40% performance with the half-baked features, etc.

This reflects the change in Silicon process economics. It used to be that with each process step you would get up to 50% price cut in the cost of transistors.

Turing stayed on the old process, so cost per transistor is relatively unchanged. But even changing process won't help much. 7nm is so expensive, that even though you can pack a lot more transistors, into each square mm, that area has shot up in cost, partially mitigating the economic benefit of the die shrink. So damned if you do, damned if you don't.

Bottom Line: Transistor prices are NOT shrinking like in the past.


Lets compare 2080 die vs 1080 die, since that is the most applicable position comparison (2080Ti die is so large, it's a yield nightmare).

1080: 7.2 Billion Transistors

2080: 13.6 Billion Transistors

13.6/7.2 = 89% more transistors.
($800/$600) = 33% more dollars, so really not that bad. ;)

Seriously though, 89% more transistors in a world where the cost per transistor isn't going down significantly changes things from the "good old days".

So holding your breath and waiting for a return to the "good old days" won't work. No supplier facing an 80-90% production cost increase of a major part is going just eat that cost, they are going to pass it on.

Expect stingy perf/$ improvements going forward as the new normal.
 
...
I don't know enough about Tensor core use cases to say whether they are good to have or not. Maybe they just haven't found the killer use case for it when it comes to games. Maybe we could have them handle AI processing or something in the future, who knows.

I am sold on raytracing tech though. Next few generations will make it a norm and hopefully with the performance increases needed too. As a 2080 Ti owner I am still excited about being able to use it right now even if it incurs a hefty performance hit. I don't mind too much as I've never been one to chase the absolute highest frame rates and I don't play competitive multiplayer shooters.

Yes I think Ray Tracing is an important step, and this was not another attack on RTX or Ray Tracing in general. I think the RT cores are doing an unmatched job on Ray Tracing. What they do can't be replicated without much great cost on GPGPU/CPU HW. RTX might have been better if it ditched Tensor cores in favor of more RT cores.

It's the Tensor cores that seem superfluous. A solution looking for a problem. They were tacked on for Ray Tracing in theory, except the job they do their (denoising) can apparently be done just as well with GPGPU/CPU code. DLSS I had high hopes for, but I have seen nearly every implementation picked apart and come up short, often worse than simple resolution scaling, which for equal performance, often has better image quality than DLSS.

DLSS requiring specific game pre-training and still coming up short compared to simple resolution scaling makes it an abject failure to deliver IMO.
 
It's the Tensor cores that seem superfluous. A solution looking for a problem. They were tacked on for Ray Tracing in theory, except the job they do their (denoising) can apparently be done just as well with GPGPU/CPU code. DLSS I had high hopes for, but I have seen nearly every implementation picked apart and come up short, often worse than simple resolution scaling, which for equal performance, often has better image quality than DLSS.

DLSS requiring specific game pre-training and still coming up short compared to simple resolution scaling makes it an abject failure to deliver IMO.

One problem with DLSS is that it usually gets analyzed when it comes released with a game, which is usually not the best implementation. We saw with Metro Exodus that they improved it over time and I've seen in my firewall logs that Nvidia does request for updates to what I believe is that stuff.
 
One problem with DLSS is that it usually gets analyzed when it comes released with a game, which is usually not the best implementation. We saw with Metro Exodus that they improved it over time and I've seen in my firewall logs that Nvidia does request for updates to what I believe is that stuff.
Why is that a problem? Periodically updating the algorithm it uses for LSS seems like obviously how it should work, to me.

I'm no expert, but I really like DLSS. Not every game supports it, but the ones that do see a pretty big performance improvement in my testing on my own system, and while some complain about the image, I can't really tell a difference at 4K. If it's the tensor cores that are being recruited to do whatever it's doing, then I'm glad they're there.

Speaking for myself, I'm also glad that they're there because I keep meaning to learn some machine learning/data science programming, and I feel like I'd eventually want to be able to use them to accelerate that, even if I'm only doing it at the hobbyist level at home. I'd obviously go get access to a farm of GPUs for any production work, but having access to the real hardware it has to run on always makes software development easier, versus doing without and extrapolating.
 
Why is that a problem? Periodically updating the algorithm it uses for LSS seems like obviously how it should work, to me.

I'm no expert, but I really like DLSS. Not every game supports it, but the ones that do see a pretty big performance improvement in my testing on my own system, and while some complain about the image, I can't really tell a difference at 4K. If it's the tensor cores that are being recruited to do whatever it's doing, then I'm glad they're there.

Speaking for myself, I'm also glad that they're there because I keep meaning to learn some machine learning/data science programming, and I feel like I'd eventually want to be able to use them to accelerate that, even if I'm only doing it at the hobbyist level at home. I'd obviously go get access to a farm of GPUs for any production work, but having access to the real hardware it has to run on always makes software development easier, versus doing without and extrapolating.


I think the problem he is referring to, is that when DLSS gets tested early, the implementation is sub-optimal, thus painting a poor perception of DLSS, when it could be better if tested later. Not agreeing, just pointing out the implication.

I think the main reason some people report it looks as good as native at 4K, is really because 4K is overkill, not that DLSS is doing anything good.

In testing where they actually go at it in detail looking at the smallest detail frame by frame, 4K (2160P) DLSS usually comes in second to simple 1800P resolution scaling that delivers about the same performance.

From everything I have seen 1800P wins pretty much across the board.

Because it just works, delivers equal performance, better image quality, has no FPS limitation, won't break if you use graphic mods, requires no network training.
 
I think the problem he is referring to, is that when DLSS gets tested early, the implementation is sub-optimal, thus painting a poor perception of DLSS, when it could be better if tested later. Not agreeing, just pointing out the implication.

I think the main reason some people report it looks as good as native at 4K, is really because 4K is overkill, not that DLSS is doing anything good.

In testing where they actually go at it in detail looking at the smallest detail frame by frame, 4K (2160P) DLSS usually comes in second to simple 1800P resolution scaling that delivers about the same performance.

From everything I have seen 1800P wins pretty much across the board.

Because it just works, delivers equal performance, better image quality, has no FPS limitation, won't break if you use graphic mods, requires no network training.
As far as DLSS goes the DPI would skew peoples perceptions on it a fair amount. Another generational change to underlying tensor cores to it could improve the quality further to the point where at 4K it's highly acceptable and at 1440p a bit like it is now at 4K being a option that's rather hit or miss.

Very true. So why would
NVidia go adding a whole bunch of them to the die that don’t really do a whole lot?
Perhaps it takes up less transistors overall relative to the performance/image quality impact?

Outside of that it's not all about gaming it's kind of the same as CUDA that doesn't "purely" benefit gaming either though in same instances can be beneficial.
 
Outside of that it's not all about gaming

Well, I was under the impression the GeForce brand ~was~ all about gaming. If you wanted CUDA or anything else, there's Titan, Quadro and Tesla that are much more suited for it, and are not branded as all about gaming.
 
  • Like
Reactions: ChadD
like this
Still raining where you guys are huh?

Over the last 20 years I admit I have almost always been a mid range card guy. I say almost, as a few times here and there when the performance gains where tangible I stepped up. In Canada where I am that means that pretty much every couple years I'm in for $250-400 Canadian and every 3rd or so purchase I have splurged and spent $500-600. That used to get me top of the line performance.

Since the mining craze I am now looking at the old high end price for a decent mid range.... and now NV thinks top of the line should be $1500 Canadian.

Here is the thing I can go to the local part shop right now... buy a 2700x a good 470 board for it from Asus or Gigabyte... drop 16gb of DDR4 3200 in it... and a 1GB M2 from Samsung. For right around a grand. Now I know 2700x isn't top of the line.... but damn NV seriously thinks that we should be paying more for a GPU then the entire cost of the rest of our systems. The cost right now of the 2080ti where I am is almost identical to what it would cost me to build a 2950x system.

I get all the reasons GPUs have went up... and I'm not QQing. If I really felt it was worth spending $1500 on a GPU I would. I have spent more on worse investments. I do think we are getting to the point of sillyness though. If game streaming takes off... it will be in part thanks to NV thinking their top of the line cards should cost more then everything else in a upper mid-high end PC combined.
 
Last edited:
I'm not mad prices went up. I don't ~have~ to spend that kind of money. And I'm not planning on it.

But I am nostalgic for the days when I didn't have to spend that kind of money to get that level of performance. That leads me to believe that these things don't ~have~ to cost that much.

And it does concern me that prices will continue to escalate. I guess that's econ 101 though - supply/demand. For those of you that are willing to afford it, more power to you.
 
Funny how video cards can make people cry.
If you think a card is not worth it...don't buy it...simple as that!
No one is forcing you....but stop the buthurt "the grapes are sour" whine....it's boring.

The market will decide...not matter the whines...it like SJW's...a small minority, being far to vocal for their size...

TL;DR: Buthurt people whine to much
 
Very true. So why would
NVidia go adding a whole bunch of them to the die that don’t really do a whole lot?

To improve performance.

The 2080 gets something like 35% performance gain over the 1080, so they are giving a respectable performance gain.

Bottom line is that the days of getting that kind of performance increase with ZERO cost increase are behind us, with or without RTX.

It looks like RTX features have become a scapegoat for changing transistor economics.
 
I'm not mad prices went up. I don't ~have~ to spend that kind of money. And I'm not planning on it.

But I am nostalgic for the days when I didn't have to spend that kind of money to get that level of performance. That leads me to believe that these things don't ~have~ to cost that much.

And it does concern me that prices will continue to escalate. I guess that's econ 101 though - supply/demand. For those of you that are willing to afford it, more power to you.

Net income is 18%, which is a little high, but not absurd. Remember the 2080ti has a 60% larger die than the Titan X Pascal but costs the same. Or 12B vs. 18.6B transistors. Even comparing it against the 1080ti the pricing isnt' absurd. I don't think they are gouging as bad as people make it out to be. Their financials don't show that, either.
 
Back
Top