Vega literally kills triangles

>>61755919
I use Nvidia :3

>>61755919
Oi vey those triangles were innocent.

Can anyone present a demonstration where 128x Tessellation shows any difference from 64x?
In everything I've seen even going from 32x to 64x was extremely minimal.

>>61755987
Seems like an exponential bump, going from 64 to 128; I don't think it'd be a subtle difference judging by those numbers.

Gaymershill can you guy fuck off

>>61755919

wouldnt this fuck up some screen space stuff?

that'd actually be proper useful for rendering fuck huge shit
I think some games already do this though

>>61755943
muh 6 gorillian triangles

>gpu literally decides not to render stuff and there is no option to turn this off
That's fucking retarded, why the fuck would I want a gpu that decides it just won't play my game, how the fuck did they even thought of that.
THE WHOLE POINT OF A GRAPHICS CARD IS TO RENDER VIDEOGAMES WHY THE FUCK DID THEY MAKE IT SO IT DIDN'T RENDER WHAT THE SHIT

>>61756227
Yep, it's going to throw away all the polys you can see instead of the ones you can't. Fucking brilliant.

>>61756227
>WAAAHH WHY DOES MY GPU REFUSE TO RENDER TRIANGLES THAT ARE NOT VISIBLE?!!!

>>61756227

z culling has existed for ages, this is a better way of doing it

File: 1485083307083.png (4MB, 1543x1615px) Image search: [Google]

4MB, 1543x1615px

>>61755919
Innovating at the same time that surviving, i wonder what AMD would do with massive resources like Intel has.

>>61755919
I don't fucking get it.
Vega has so many great features over GCN4 that should make it perform so much better in games than it currently does, but instead it's worse than Fiji clock for clock while drawing more power.
On the other hand, it demolishes everything else in workstation performance.

So why the fuck does it have such a piss poor gaming performance?

>>61756603
We'll see if this is working or not when RX Vega launches.

>>61756603
drivers maybe

>>61756159
>rendering

>>61756227
You're a fucking idiot

>>61756603
because it's a workstation card

>>61756603
Because FE doesn't have this feature enabled, nor several others. It's being added in the driver update for RX's launch.

>>61756603
Retard

It's a workstation care performer because that's what it is

>>61756227
> WHY THE FUCK DID THEY MAKE IT SO IT DIDN'T RENDER WHAT THE SHIT
'Cuz devs are retards

>>61756629
>>61756813
>Muh warkstorton card xD
And yet Quadro cards get the exact same gaming performance as Geforce GPUs at the same clock speed.

>>61756728
This seems kind of implausible, we've been seeing results like FE, both leaked and official, for at least half a year now, hell I think it started in 2016.

Have AMD really never fucking tested these features in any single game? It seems so implausible, what the fuck are they doing? Why are they even showing multiple official demos of Vega running games if they don't have essential features working? They're going to get all these features working in like a single month when we haven't even seen a single controlled demo of them working this late in the game?

>>61756915
>I don't know how driver development happens
Honestly your post is barely worth responding to when its clear you're so grossly stupid. All of the tech press knows that Vega cards in the hands of the public know are using beta drivers without certain features enabled. This isn't a mystery.

>>61755919
Rasterization based rendering has no productive use. Gaming is a time and money sink and subtracts value from the global economy as a whole. No non of that performance matters in the grand scale of things.

>>61756959
>muh beta drivers
If you use Vega you're a beta alright

File: Z1Wmcqk.jpg (61KB, 700x523px) Image search: [Google]

61KB, 700x523px

>>61757005
This wasn't even half as clever as you thought it was.

>>61755919
What is Steve's background (the GN editor?) Engineering?

>>61755919
Wouldn't this screw shit up for reflections or no?

>>61756585
>i wonder what AMD would do with massive resources like Intel has.

Suffer from hubris for being top dog and do shady shit to the consumer while cranking out inferior product. Like Intel currently is.

It's just the nature of monopoly.

>>61757340
if they're using really shitty reflection calculations, maybe.

Did none of you fuckwits even notice it says "Under embargo until June 29, 2016 at 9 a.m EST."

>2016
TWENTY FUCKING SIXTEEN.

>>61755919
yeah its interesting i saw the interview seems like they found a way to cull literally everything if they so wish without a perfomance hit

>>61755987
if its nvidia game probably they have an underground ocean of triangles like crysis 2 LOL

>>61756614

>MSAA
>Rendering
>Zero area triangles

Maybe if you've got shortbus high school students that are making these scenes badly.

>>61756997
Neither does watching sports and so many other things

>>61755919
Better test Vega with AA turned off.

File: city-trees-water-mesh-620.jpg (291KB, 620x496px) Image search: [Google]

291KB, 620x496px

>>61757888
this is nvidia's way of fucking up amd cards

>>61755919
Where did this image come from? I didn't see this one in the launch slide deck on TPU, the only one with graphs and culling was that one with "Vega fastpath" that was like 10% faster than "native pipeline"

>>61757924
this is from polaris the same engine is on vega but a lot bigger

File: siggraph_vega_architecture_47.jpg (73KB, 1024x576px) Image search: [Google]

73KB, 1024x576px

150W Nano when?

I feel like this is a thing game engines should be doing, not having to rely on the gpu

File: barrier-dx11-mesh[1].jpg (76KB, 620x600px) Image search: [Google]

76KB, 620x600px

>>61757914
That's nothing compared to the concrete barriers.

>>61756227
>prey animals, have their two eyes positioned on opposite sides of their heads to give the widest possible field of view, panics when vision is constricted to only the front

>>61758059
you would let a dev that release a game that has more bugs than hours to play to actually be able to tell the gpu what to cull and not?
especially when you have to deal with shitworks?

>>61756611
>>61756613
>>61756728
Are most of the new features driver dependent or do the game engines need optimization for them? Does this triangle culling happen automagically without developers targeting a specific api?

>>61758073
In the current climate of extremely lazy Devs, no

>>61758002
>admitting they're still behind pascal

>>61756603
It's funny because the features would indicate RX Vega should surpass the 1080 Ti with these features 'enabled' but even AMD are advertising it as a 1080 competitor, clearly the features don't do much at all

>>61758310
It just means Vega is clocked to hell at stock and does fine underclocked, meaning Raven Ridge will be stellar since its clocks are gonna be low.

>>61758335
It means vega is just like the Fury series, the Nano is the highest binned consumer card while the 56/64 are just overvolted to shit because they're bad chips.

https://devblogs.nvidia.com/parallelforall/inside-volta/

>The new Volta SM is 50% more energy efficient than the previous generation Pascal design, enabling major boosts in FP32 and FP64 performance in the same power envelope.

THANK YOU BASED NVIDIA

>>61758358
how will AMD keep up they just barely caught up with Kepler?

>>61758002
I hope it's priced the same as a Vega 64. I was tempted for an R9 Nano but it had only 4GB vram back then.

>>61758358
Volta SM is 50% more energy efficient

that's what i like to hear.
AMD needs to up its efficiency also.

>>61758068
So this is the way it is meant to be payed?

>>61758070
lol

>>61758358
>Data Center GPU
hwhat?

>>61758358
>>The new Volta SM is 50% more energy efficient than the previous generation Pascal design, enabling major boosts in FP32 and FP64 performance in the same power envelope.

implying that nvidia will enable the full cuda support to gaming cards

keep dreaming

>>61759004
its literally not but ok

you see boris nvidia failed to mention that this 50% on the tensor cores comes without them using the A.I ISA literally having the sm sitting there doing nothing
its like amd saying we manage to achieve 50% more efficiency on the HWS

>>61757888
>zero area triangles
That is THE way to join models in same indexing mode. OpenGL ES 2.0 guide book specifically recommends it.

>>61757914
>this kills the amd

>>61755919

But, how many buttcoins can it manufacture per shekelsecond? Let's keep perspective on our priorities here. Miners are people, too, you know.

>>61759454
good luck pulling shit like this on vega tho

>>61759410
GV100 is also clocked lower, this is probably where all the efficiency comes from.

>>61759508
the real shit will come once nvidia installs a hardware sc back to their cards and their tdp will skyrocket

>>61756227
this is just bait right? nobody is this stupid right?

Imagine you are in the hospital and they are using AMD to do your medical imaging and the GPU just decides not to render your tumor.

>>61756211
yeah they just do this to gimp AMD cards lel

>>61756227
Crysis 2

>>61759528
But that's wrong

Volta already has built-in independent thread scheduling and yet it's 50% more power efficient than Pascal

>Volta GV100 is the first GPU to support independent thread scheduling, which enables finer-grain synchronization and cooperation between parallel threads in a program.

https://www.hpcwire.com/2017/05/10/nvidias-mammoth-volta-gpu-aims-high-ai-hpc/

>“It’s actually pretty remarkable to me that we were able to get more flexibility and better performance-per-watt. Because I was really concerned when I heard that they were going to change the Volta thread scheduler that it was going to give up performance-per-watt, because the reason that the old one wasn’t as flexible is you get a lot of energy efficiency by ganging up threads together and having the capability to let the threads be more independent then makes me worried that performance-per-watt is going to be worse, but actually it got better, so that’s pretty exciting.”

AYYMD is getting destroyed in power efficiency

>>61755919
Vega literally kills Radeon market share

>>61760119
Vega will probably be the best performing card on the first level of Crysis 2 by far now.

>>61755919
like all those power point slides - "magic technology no one even thought about, we did it" - 2% increase in best case in games

File: Vega_FE_Memory_Bandwidth.png (129KB, 800x574px) Image search: [Google]

129KB, 800x574px

>>61756603
Because it APPEARS to have 20% less raw memory bandwidth than Fiji, whether it be to drivers being fucked, sandbagging or a hardware flaw (unlikely, they had time to catch and respin already).

You tell me how much of this 20% regression in memory bandwidth will get fixed and I'll tell you how much Vega's performance will improve in games and ETH mining.

>>61761278
Why.
What the fuck.

File: b3d_fillrates.png (62KB, 1016x1203px) Image search: [Google]

62KB, 1016x1203px

>>61757005
Yeah, a 25% performance regression in texture fill-rate clock-for-clock vs. Fiji is TOTALLY logical, right?

>>61761278
fiji has double the hbm stacks idiot

>>61757731
>He doesn't know that AMD's primitive discard acceleration debuted on Polaris and wasn't used at all on Fiji.

>>61761323
HBM2 stack has double the bandwith per pin though.

File: 900x900px-LL-689b3a99_6.png (124KB, 900x775px) Image search: [Google]

124KB, 900x775px

>>61758331
>It's funny because the features would indicate RX Vega should surpass the 1080 Ti with these features 'enabled' but even AMD are advertising it as a 1080 competitor, clearly the features don't do much at all
Gee, I wonder if a seeming ~30% regression in effective texture bandwidth vs. Fiji might explain that?

Wait somebody with half a brain did B3D suite benching?

>>61761319
wrong post, m8

>>61761374
>Wait somebody with half a brain did B3D suite benching?
I found the benchmarks into my own digging into this, they're from a German site.

This has me pulling my hair out because it seems logically impossible that you would design a higher clocked version of Fiji to have LESS memory bandwidth, and yet there appears to be a massive memory bandwidth regression. You would also expect a large bandwidth savings from primitive discard and DSBR, yet RX Vega 64 appears barely faster than Vega FE.

So either the drivers are/were totally fucked or RTG is sandbagging like crazy, or it's some of column A and some of column B.

>>61761453
>So either the drivers are/were totally fucked or RTG is sandbagging like crazy
It's both.
Pro drivers are ready though, they actively demoed WX9100 and SSG in their SIGGRAPH booth.

>>61758096
I dunno about vega but drivers for AMD over the last like 2 years have been targeting specific software (games)

File: mining.webm (498KB, 480x270px) Image search: [Google]

498KB, 480x270px

>>61755919
Vega kills everything, including your power bill.

Based Raja.

>>61757914
>>61758068
an entire FUCKING SEA under each level

>>61761254
>magic technology no one even thought about, we did it

that's not what this video is about

not rendering non-visible scenes is an old idea

>>61761453
Either Vega's memory controller has the same issue that effected Fiji's effective vs peak bandwidth, or current drivers were so cobbled together that things are only in a barely working state.

>>61761713
As i said, WX9100 and SSG are working nice and dandy.

>>61761713
considering the hardware side of things is held together with paste and popsicle sticks I wouldn't be surprised if the software side was similar

>>61756188

Normal shaders can't just write to random (sub-)pixels. UAV shaders can, but that's not really relevant here.

>>61758237
The devs aren't lazy, they are cheap. NVIDIA's middleware is a good deal, that it's to designed to fuck AMD architecture and allows nearly no way to remedy that for the game developer ... oh well, it's free. Win some, lose some.

>>61761612
Yes, this is how CryEngine works ever since first FarCry

>>61760139
They're talking about the "muh big die" version for AI research that costs almost as much as a small house. How much of that ends up in the consumer product out next year remains to be seen.

>>61755919
vega will compete with the gtx 1160

>>61761967
That's horseshit unless you're retarded and expect 60%+ performance from GV102 without a dieshrink

>>61761992
No one expects that besides kids & shitposters.

>>61761992
That would be real nice but very improbable, GP102>GV102 is gonna be 40% tops in whatever scenario Volta is good at.

>>61762028
>40%
Unlikely.
It's almost certainly 20% bigger dies with overall 25% perf improvement.

>>61762055
I said tops, I'm thinking of an average of 30%, but it all depends on how big the dies will move compared to Pascal

>>61762072
>yfw Nvidia and AMD will never reach the ultimate kikery that is Intel, charging $350 for 100mm^2 of BULK silicon

>>61762072
Consumer Volta is 20% bigger tops. Otherwise the shareholders will choke JHH.
>>61762081
Absolutely nothing will top EPYC at margins per top SKU.

File: 538.jpg (286KB, 1841x705px) Image search: [Google]

286KB, 1841x705px

>>61762028
Nvidia just need to stop selling mid tier GPU's at high end prices and they will increase in performance between generations.

The 600 series did not age well, 1080's will have the same trajectory into obscurity as more is asked of it. Meanwhile the AMD 200 series still had plenty of fight left in it.

>>61762140
Nvidia's life depends on margins on consumer GPUs.
AMD has CPUs.
Actually fucking GOOD CPUs.

>>61762090
Consumer prices, EPYC margins are pretty fine but they take into account support and validation.

Consumer bulk silicon that's out of the vendor's hands doesn't include these things which are more difficult to calculate

>>61756959
No, that's the thing. I do know how software development works, you seem to be woefully ignorant of that, but just clever enough to hurl insults. Working, public release-ready software does not pop out of the ground when the magic release date rolls around. Software development is a gradual process which requires extensive testing, so it is exceedingly strange that this late in the game AMD's software support for Vega seems to be the same as it was half a year ago. At this point in time they should definitely have the features working, even if they aren't ready for a public release, they should at the very least be functional enough for limited demos or we should've seen a leak.

I'm not doubting that shit is disabled, what I'm doubting is AMD's claim that everything will be functional on release in like a month. For that to be plausible they would have to be hiding true performance intentionally while raping the Vega brand with a very bad reputation pre-release.

>>61762303
They heavily focused on Pro stuff to show both SSG and WX9100 at their best at SIGGRAPH.

>>61760139
Hmm, so they just improved their software based hardware scheduler.

Does this mean Volta can actually take advantage of Async features in games?

>>61762340
that's because they had no presence before in pro market
having amd support for two AAA games before they are released is quite big, specially if you consider which companies do the games
Ubi been nvidia buddy for a long long time
then there is a matter of a freaking sweeney showing up at siggraph
AMD clawing back
we simply do not know anything about RX Vega to say for sure and Pro stuff was on the surface

will this new culling work with old games though? Are drivers enough?

>>61762536
>will this new culling work with old games though?
Yes, but devs can probably explicitly target it.
>Are drivers enough?
Yes.
But they need to have DSBR working, since prim shaders are heavily reliant on it,

>>61756603
Something very early in the pipe is borked. Maybe scheduling or memory. It exhibits the same performance in openCL benches as it does in games meaning the problem is before it hits the rendering pipline (geometry processing, culling, etc)

The NCU themselves have 15% higher perf/hz compared to GCN4 and it's evident in certain shader heavy benches that hit only the ALUs hard.

>>61761347
And it's clocked about 20% lower than HBM should be (800~MHz vs 1000~MHz)

>>61762591
That's for Vega56.
FE stacks clock at 945mhz.

>>61762573
Memory. Read the thread.

>>61762573
its weird how nobody still know what it is, Fe been a long time out and nobody did proper analysis

Also once again I want to point out AMD didn't say anywhere what actual TDP of the vega is. At least I can't find it anywhere, specifically from AMD side. Undervolting drops power significantly, so if proper voltage control is in place who knows how much it draws.
A thing with ryzen power efficiency, it goes down the drain if you disable c-states in BIOS, might be same deal with vega.

>>61762629
>Fe been a long time out and nobody did proper analysis
G*rmans did.

You people are blowing Vega out of proportion, it's another Hawaii and that's it, only this time it can actually OC decently if you don't care about power usage.


Vega 56 stock runs around 1360Mhz average yet you can overclock Vega over 1700MHz, RX V56 shouldn't be far behind a 1080 anyway.

For $400(+20-50$ for AIB models) it will do just fucking fine with a heavy OC

>>61762697
Hawaii was Tahiti with moar ALUs.
Vega is drastically different from previous GCN iterations.

>>61762710
I meant in what it represents and its position in the market.

>>61762697
Source on the clocks?

>>61756603
Outdated apis. Nothing will change in gaming until Vulkan is mainstream

>>61762783
Testing on FE shows that it averages over 1400MHz (power limit), V56 has a shittier blower cooler used in Polaris and it's obviously worse than the FE cooler.

>>61762697
>mfw 400W oven for $400

$1/watt, AMD is a genius, puny Nvidia only has half.

>>61762971
UMA

>>61755987
In reality, you won't see a difference from 16x tessellation and up.

>>61763384
there is even no difference between max tesselation and no tesselation in some games. gta 5 for example.

File: 1500716056988.jpg (138KB, 1200x675px) Image search: [Google]

138KB, 1200x675px

>>61755987
Shut up goy

>>61756603
>So why the fuck does it have such a piss poor gaming performance?
Drivers.
Games are optimized to squeeze out performance out of nVidia cards, AMD is left out.

>>61762573
Memory bandwidth currently has a 20% regression from Fiji as of Vega FE, ~300 GB/s in Vega FE vs ~360 GB/s in Fiji.

>>61762697
How would Polaris perform if you got rid of all primitive culling? You are saying primitive culling AND tile based rasterization are worth ~0% on Vega?

>>61761953
There's virtually no reason for them to break compatibility and omit the most important features from consumer products. They've shown they can and will artificially limit non-gaming performance enhancements to keep the pro products thriving.
Better DX12/Vulkan performance is critical to their products mindshare going forward, Nvidia isn't going to arbitrarily pass on architectural features that will provide that better performance.

>>61764685
Taking a wait-and-see approach to their claims of independent thread scheduling and reduced power draw at the same time.

>>61763486
you have already seen the rx vega benchmarks?
SHIT ANON GIVE ME THE LOTTON NUMBERS