AdoredTV attempts to justify his position on Pascal

[OneMoar]

ill say this for the first time in ages
Nvidia has some serious work todo on the driver side

 

[alucasa]

Why reinvent the wheel?

Comes to my mind.

 

[alucasa]

I am not defending neither AMD nor Nvidia. Just saying that coming up with an entire new arch is much easier to be said than done.

 

[GreiverBlade]

Ive always thought that Pascal is an improved Maxwell, a die shrink with added refinements within (but not a completely redesigned architecture ala G70 to G80 or Fermi to kepler). Also, ROPs and TMUs are different for 1080 and 980Ti and these 2 components are pretty critical even now as it was before for standard rendering and rasterization/anti-aliasing at high resolutions.

Anandtech's (late but great and detailed review of the GTX 1080 and 1070) discusses refinements added to Pascal vs Maxwell
http://www.anandtech.com/show/10325/the-nvidia-geforce-gtx-1080-and-1070-founders-edition-review

thanks for that article (currently reading it)
if there is other improvement/refinement, it's not only a die shrink of the same GPU... ( also you point ROP and TMU ... well the one with more should outperform the one with less at same clock ... but ... it does not )

otherwise the 1080 is what? a 980-1/2 ... since in terms of shader/ROP/TMU etc etc it's between the 980 and 980Ti ... GP104 is not a cut down GM200 afaik

yep that article sum it up (and they compare GP104 to GM204 and not GM200 )
"Looking at an architecture diagram for GP104, Pascal ends up looking a lot like Maxwell, and this is not by chance. After making more radical changes to their architecture with Maxwell, for Pascal NVIDIA is taking a bit of a breather. Which is not to say that Pascal is Maxwell on 16nm – this is very much a major feature update – but when it comes to discussing the core SM architecture itself, there is significant common ground with Maxwell."

to me the main point is that :

noticed something else ... the 1070 clocking is not far from the 980Ti and sports the same GDDR5 but has way less transistor/core/TMU/ROP/bandwidth ... only a shrink and +300mhz/-68mhz core/mem make that happen? i don't think so. for a underclocked 1080 i understand it come near a 980Ti ... but if even the 1070 does it at stock...

how is it possible then, that a gpu with lower specs and a slightly higher clock ( +300mhz/-68mhz core/mem) is totally "on par" with the one with higher specs if pascal is only a die shrink of maxwell, improved indeed refinement (major...) same GPU? absolutely not.

ill say this for the first time in ages
Nvidia has some serious work todo on the driver side

pinpointed... (and they need to do that since... ages .... i had more driver issues with Nv than ATI/AMD)

because the engine that powers AMD's cart is garbage so they need to invent a better wheel to make it work

failed... the RX480 is not so bad... with new API (but that's a hard bet to do ... from AMD )

 

[ViperXTR]

( also you point ROP and TMU ... well the one with more should outperform the one with less at same clock ... but ... it does not )

Yeah, and when you put them at the same clock, FLOP also changes thus still won't be an accurate comparison.
If the architecture is more like in Fermi days where the ROP and TMU clock domain is separated from the shader domain, maybe you could put up a more accurate comparison

 

[xkm1948]

https://www.techpowerup.com/forums/posts/3495100/

Great video as now we have multiple independent source proving the same hypothesis that Pascal is not a major change in design. As I have argued in the the Volta thread this is the best way to ensure no suffering of gaming performance during the transition to Async Compute heavy DX12/Vulkan. Now it is pretty clear why Nvidia is rushing to push out the Volta. They need a complete refresh of design to keep up with the DX12/Vulkan requirement.

By the time Volta taps out Nvidia has already mastered designing on 16nm FinFET. I see it very likely Volta will be a monster in ALU size as well as frequency. Then Nvidia can take 2~3 generations of time to polish that structure to the max. All in all as an enterprise Nvidia's decision making is top notch and their hardware rarely fails to deliver.

 

[ViperXTR]

Nvidia is rushing to push out the Volta.

They need to, otherwise the US Government won't be happy as they have a contractual agreement with nVidia for Volta to be released in 2017 for their Summit supercomputer

http://www.anandtech.com/show/8727/nvidia-ibm-supercomputers

 

[brian111]

The fault with the video is in how TFLOPs are matched. Of course GPU designs with similar TFLOP performance will be nearly equal.. what do you think TFLOP measures? Then, where 1080 gets a bit ahead is entirely due to the memory compression. Seems like just a bit of false science to me. If someone didn't know that a FLOP is a way of measuring computational performance, the video could mislead them. Seems like that was the point in the video... to mislead people so that the poster's opinion is justified.

"Hey's let's match VGAs to the same computational performance, and then tell people that they are the same because of that!"

SMH.




So, me being the reviewer, I'm gonna call BS on this video, because that's just my character. I became a reviewer because I wanted to see an end to BS like this, in fact.

The number of shaders can't be matched by matching TFLOP performance. That simply matches performance in general, not shader counts. OMFG! I can't believe you're posting on here like this is anything worthwhile for anyone to watch or pay attention to. It's outright BS.

But as a point of discussion, sure, it's interesting to see how many people didn't catch this part of it. I purposely waited a bit before posting this to give some time for people to see it and form their own opinions, and to see if anyone else would notice this obvious under-handed way of trying to prove a point that amounts to spreading shit all over the internet.

So are you saying it is just an issue with the methodology or he's being inherently dishonest ("obvious under-handed way")?

 

[cadaveca]

So are you saying it is just an issue with the methodology or he's being inherently dishonest ("obvious under-handed way")?

I honestly dunno, you decide. He either knows exactly what he's talking about, or knows very little and made a mistake in his attempt to compare the cards.

Dunno why someone would mention TFLOPs and then use that as a point of compare if they knew what they were talking about, since FLOPs is literally a way to compare performance (FLOPs = floating point operations per second). So if you adjust one card to match performance of the other, and expect different results, you just might be crazy, or you simply might not know what you're talking about. Either way, it makes the video misleading, whether intentional or not.

 

[RealNeil]

the gf can't understand us after we've had a few drinks

LOL!

I can hear it now, "that's it! I'm going to bed, goodnight!"

 

[Adored]

I honestly dunno, you decide. He either knows exactly what he's talking about, or knows very little and made a mistake in his attempt to compare the cards.

Dunno why someone would mention TFLOPs and then use that as a point of compare if they knew what they were talking about, since FLOPs is literally a way to compare performance (FLOPs = floating point operations per second). So if you adjust one card to match performance of the other, and expect different results, you just might be crazy, or you simply might not know what you're talking about. Either way, it makes the video misleading, whether intentional or not.

What happens when you compare the 5.3 TFLOPS of the 780 Ti and 5.0 TFLOPS of the 980? Now increase the 980's TFLOPS by 0.3 and retest. Do you think maybe you're the one that doesn't know what you're talking about here?

 

[OneMoar]

What happens when you compare the 5.3 TFLOPS of the 780 Ti and 5.0 TFLOPS of the 980? Now increase the 980's TFLOPS by 0.3 and retest. Do you think maybe you're the one that doesn't know what you're talking about here?

nope that is not how it works you can't just scale clocks to 'match the number` thats not how it works there are other factors at play here memory compression for one ... preemption and better dynamic balancing
what you did was this

you took two engines revved them both to 4 thousand rpm and decided they where the same because they made the same power at that given RPM (so by your logic at V8 that makes 300HP @ 4KRPM is just as good as a turbo 4 that makes 300 @4K RPM and 650 @ 9k)

frequency does not always denote a liner change that is your fatal assumption

and anybody here will tell you that dave knows his stuff and I would wager a large chuck of cash that he could teach you a thing or three

 

[Adored]

nope that is not how it works you can't just scale clocks to 'match the number` thats not how it works there are other factors at play here memory compression for one ... preemption and better dynamic balancing

All of which should improve the performance on the 1080, not the 980 Ti. In the case of the 980 vs 780 Ti the numbers are so overwhelmingly in favour of the 780 Ti yet it still loses on a flop vs flop level. Why? Because Maxwell is a vastly superior architecture to Kepler.

frequency does not always denote a liner change that is your fatal assumption

and anybody here will tell you that dave knows his stuff and I would wager a large chuck of cash that he could teach you a thing or three

There is only a 10% difference between 2816 and 2560 shaders. If the lesser GPU has 10% higher clock speeds then numbers will be very close. Shaders also do not always denote a linear change, which is an assumption that many people falsely make. The overall differences in this case are tiny and on the same architecture you can expect the same performance when most other aspects (ROPS, bandwidth, memory amounts) are not an issue. That was another reason for downclocking the 1080 by a lot instead of overclocking the 980 Ti.

That's what the results showed. Dave is welcome to try to teach me a thing or three, let him talk for himself if that's what he wants to try.

 

[RejZoR]

Easiest way to compare would be if you could in software limit the number of shaders graphic card can use to process stuff. So you would just limit both cards to same shader count and same clocks and you'd compare architectures head to head.

 

[the54thvoid]

I'm still not watching video as the Nordicised Scottish accent hurts my patriotic sensibilities, 'ye ken?'.

But, from all the arguing, I can say Cadaceva does know his stuff. He also isn't a diplomat, mostly, so perhaps before you two start cock waving, perhaps ease off on them zippers.

As for generics, it's well established, Pascal is an evolution of Maxwell, a very small evolution. Volta will be different. Likewise, Fiji was a small evolution on Hawaii. AMD's move to GCN from VLIW was their last large jump. Likewise, I'd say Fermi to Maxwell via Kepler was the last real Nvidia transition.

As for you tube and bloggers in general, I never trust them fully. I work in fitness and all I've learned from YouTube and blogging is, the real experts don't use social media. And that's the lesson, @Adored has stated its an interest of his and Dave's our mobo reviewer. You'd need an accredited hardware designer (like my brother) to wade in and say "stop talking projected conjecture based on poor assumptions, here's the cold soldered science".

Pah.

 

[Vayra86]

Basically the dude makes enjoyable videos for nerds, and in a previous video that came after the announcement of the 1080 and 1070, he noted that the new architecture hadn't changed much in the node shrink, it was just a more densely packed maxwell with higher clocks.

Apparently he got a lot of flack for it and decided to test it.

The results are interesting and in line with the speculation.

It doesn't mean much, it was just an interesting video for anyone who geeks out about this stuff.

Yup. Same as I've been saying ever since those high clocks and the per/clock performance was getting clearer. All Nvidia did was streamline the SMX to allow for higher clocks, and FinFet does the rest, the new voltage circuitry creates the efficiency jump from Maxwell > Pascal together with the smaller node, the arch is just copy paste and the circuitry changes only serve to push 'Maxwell v2' closer to its max clock out of the box. Anyone remember me saying 'Nvidia ate up our OC headroom'? There you go.

This conclusion doesn't require a rocket scientist either, just compare clock differences and the amount of shaders compared to past gen. It was easy to see - no need to jump on Youtube for that. Just read specs and see benches.

 

[laszlo]

watched the video and read comments here also

think dave has a valid point as you can't compare cards like this;every small arch. change,improvement from last gen., has an impact which comparison didn't show the way is made by limiting certain freqv. ; i'will like to see a comparative crunch computation and execution time.

srry adored but i can compare your way of testing with limiting a bugatti to 500 hp & 200 mph and comparing it with a ferrari with same hp &speed limit; maybe both will reach max speed in same time;this doesn't mean they have same tech. characteristics

 

[OneMoar]

watched the video and read comments here also

think dave has a valid point as you can't compare cards like this;every small arch. change,improvement from last gen., has an impact which comparison didn't show the way is made by limiting certain freqv. ; i'will like to see a comparative crunch computation and execution time.

srry adored but i can compare your way of testing with limiting a bugatti to 500 hp & 200 mph and comparing it with a ferrari with same hp &speed limit; maybe both will reach max speed in same time;this doesn't mean they have same tech. characteristics

butt maaaaaaaaaaa they both have 4 tires they are the sameee they both go vroomvroomv

 

[revin]

So what about a highly oc 980/ti vs the 1080, just what happens there?
[Edit: on LN2 ect.. that kind of OC to get closer to the 2Ghz ]

A quick Google of images show a better 980ti FS Extreme score over the 1080.

What I'm wanting to understand in all this, is what happens when you super tune the 980 instead of detune the 1080..................................
Does it need to show both ends of the scale to balance ??

 

[OneMoar]

So what about a highly oc 980/ti vs the 1080, just what happens there?

A quick Google of images show a better 980ti FS Extreme score over the 1080.
What I'm wanting to understand in all this, is what happens when you super tune the 980 instead of detune the 1080..................................
Does it need to show both ends of the scale to balance ??

if you super-tune the 980 you get a 1060

 

[R-T-B]

Why? Because Maxwell is a vastly superior architecture to Kepler.

For gaming yes. For fp32 compute, not so much.

 

[RejZoR]

So what about a highly oc 980/ti vs the 1080, just what happens there?
[Edit: on LN2 ect.. that kind of OC to get closer to the 2Ghz ]

A quick Google of images show a better 980ti FS Extreme score over the 1080.

What I'm wanting to understand in all this, is what happens when you super tune the 980 instead of detune the 1080..................................
Does it need to show both ends of the scale to balance ??

Maybe because it's easier and more reasonable to slow down the faster one than superclock the slow one. For clearly obvious reasons. Downclocking is easy. Superclocking is difficult even with a lot of experience and right equipment. And even then you can hit annoying issues like with RX 480 which has a cold bug and starts to go haywire (PCIe drops into power saving mode) if you cool it below zero °C. So, going the opposite direction is more reasonable and logical approach.

 

[cadaveca]

What happens when you compare the 5.3 TFLOPS of the 780 Ti and 5.0 TFLOPS of the 980? Now increase the 980's TFLOPS by 0.3 and retest. Do you think maybe you're the one that doesn't know what you're talking about here?

As I said earlier, you're not entirely wrong in your synopsis; it's how you go about trying to prove your idea that is wrong. 780TI and 980 are wildly different GPU designs on the same node. 780 TI is 2880 shaders (250W on 28nm) and 980 is 2048 (165W on 28nm). How is it possible the 980 beats the 780TI? Explain that in an exact way.

So slow down a moment. As others have said, I am not always tactful on the forum because well, that's just me. The whole thread is a bit stinky to me... I surmise you personally know the OP, too. So ignore my attitude (which you have done for the most part, much respect to you for that) and keep discussing.


Do also keep in mind I'm not trying to educate you on anything. I do know a bit about hardware, but I am not TPU's GPU specialist. I do motherboard, memory, CPU and full PC reviews here. However, it's up to you to refute what I say exactly, since I'm already established here (as you've seen by other poster's comments) and you are not. I'm not saying that you know less than I do, or that my position here truly means anything; that's just how society works. So please explain further. Maybe I will learn something. I'm just a hard person to teach because I am overly confident.

 

[yogurt_21]

interesting arguments here. For one you have FLOPS as a measurement of performance on the other you have some HP analogies.

the funny thing is if you combine those two you get a different story entirely.

If you take a chevy, a ford, a toyota, a bmw, ferarri, porche, bugatti, suburu, and a mercedes and bring them all to the same HP rating at the same rpm, would you really expect them all to have the same performance?

Ie the same 0-60, the same quarter mile, the same times around the nurnburg ring, the same time on a dirt track, the same finish in the Datona 500, etc...

No you certainly would not. Not at all. And that's not even taking into account model and design from each of those manufacturers ie trucks, performance cars, family sedans, etc. Plus within each of those categories you have more specific differences.

So with all due respect to Dave, I feel his argument is about as weak as it gets. After all he IS a mobo reviewer, doesn't he take a look at boards with the exact same chipsets? Or do features not matter? because last I checked board prices for a specific chipset have several hundred dollars between lowend and highend for that same chipset.

architectures do better in different environments using different features, shading, tessalation, vr, 3d vision, FOV, Lighting effects, etc.

so I can't imagine why you would have a beef of making the FLOPS identical and running a few benches to see how each handles those different features. It would be no different than pretty much any comparison Top Gear (with the big 3, Clarkson, Hammond, and May) has done time and time again. They got 3 nearly identical spec vehicles and put them through their paces expecting different results. Were they crazy? well maybe they were but not for that reason.

Will the same performance make it the exact same arch? Of course not. But will it be interesting and entertaining? F yes!

 

[OneMoar]

I am of the opinion that entertainment has no place in a technical review
there is a difference in being good on camera and then just going out of your way to prove that you are 'right` by bending the numbers and creating abritrary cases for your pov

thats how we end up with trash like LTT
no thank you sir ill keep my 'cold soldered science` as @the54thvoid put it

meanwhile AMD has been recycling GCN for the past 4 years and calling it a new architecture every-time but I don't see anybody taking a mesuring stick to them