AMD Polaris Thread: Radeon RX 480, RX 470 & RX 460 launching June 29th

[MTDEW]

If he is doing this live can't someone post on his video and ask for the memory clocks?

He's been asked several times and he seems to just ignore the request / doesn't even acknowledge it.....no idea why.

 

[imported_bman]

Comparing the two benchmarks the 480 performed around ~82% of the r9 390 system in GTA 5. R9 390 system also has motion blur enabled and AMD CHS soft shadows, advanced graphics might have some differences also. I know AMD CHS has around a 10% performance hit and motion blur is practically free. So the 480 is probably around ~75% of the 390 in GTA 5 on the same settings. Systems difference will also cause some differences, not sure how much.

RX 480 8GB
i5-6400
2x4GB 2133MHz DDR4
1TB 7200RPM HDD

vs

i5-4690
msi r9 390 gaming 8g
asus h81-gamer
kingston 8gb(2x4gb) 1600mhz cl9 ddr3
samsung 850 evo
corsair cx750

 

[William Gaatjes]

GCN ROPs are decoupled. 7970 have 32 ROPs with 384bit bus. Limitation to number is only by AMD choice.

Aha. I see.

edit:
Comparing all architectures.
It has always been a multiple of 8 for the later faster generations.
8, 16 ,32, 56 or 64.
There is some design rule of thumb there.

 

[Headfoot]

I am trying to understand this, so do not hold it against me when i am totally wrong.
I read in the gcn white paper from AMD that the scalar unit is mostly for control flow and address calculation and generation and not actually graphics or shader calculations.

Are the scalar unit instructions that control the flow of instructions for the vector units not better suited to be executed out of order ? I can imagine that doing pipelined arithmetic instructions in the vector unit would indeed be difficult to execute out of order. Because i assume, the instructions would be very dependent on the previous results from previous instructions would. That would mean a lot of instructions need to be tracked before independence would be discovered and that may reduce efficiency. Something a compiler may be better at.

That would agree with what you write.

Yeah I wouldn't take those slides to mean OoOE. That's a somewhat specific thing but at the same time it can mean a lot of different implementations.

Key to the OoOE concept is that you still get back the same result as if you had done the operation in a truly blocking based serial manner, except (transparently) it is not actually truly serial.

Here, they are exposing a different feature. They are allowing you to forgo the guarantee of the same result from the hardware/driver as an in-order operation. As it explains, this is not a problem where you as the developer know that in-order or out-of-order results for your particular command will be identical. But if you ran a command that did need to be completed in order you might get race conditions or an incorrect result, whereas with OoOE you would not.

It's a related but distinct concept.

 

[William Gaatjes]

I have been wondering about the rumor that GCN 4 has Out of order execution.
And after reading, it could be that the scalar unit might be OOOE.
The vector units are in order execution and i think it would be very power hungry if not. In order is more energy efficient. Even at small processes.

But i think the out of order execution that is referred to when speaking of GCN is nothing more then the ability for true chip level asynchronous compute by courtesy of the ACE.
We will have to see until we get more information after 6 days.

 

[Mopetar]

Are you buying ROPs or a video card?

If only having 32 ROPs is a performance bottleneck is certainly matters. 64 may well be redundant or overkill, but if the real number of ROPs for balanced performance needed to be ~40, then 32 is insufficient.

 

[desprado]

Guys, I had some free time this morning so I went ahead and did a comparison video. While our specs are not the same, I wanted to see the difference.

Overclocked GTX 970 1522/8000 vs. his RX 480 @ 1266/8000 ? in GTA V.

The settings are as close as I could get them. Since my card is limited to
4GB/3.5GB :\ I couldn't max out the Extended distance scaling. I'm not sure how much difference it makes?

Anyway, here is a video of the comparison.

https://youtu.be/T0b-lBEkfjg

AMD will never achieve OC like Maxwell even Pascal cannot.

GTX 970 is still a better card and glad to know that. I fell sorry for those who sold their R9 390 or X for 480. it is downgrade in term of performance.

 

[Headfoot]

I have been wondering about the rumor that GCN 4 has Out of order execution.
And after reading, it could be that the scalar unit might be OOOE.
The vector units are in order execution and i think it would be very power hungry if not. In order is more energy efficient. Even at small processes.

But i think the out of order execution that is referred to when speaking of GCN is nothing more then the ability for true chip level asynchronous compute by courtesy of the ACE.
We will have to see until we get more information after 6 days.

Im not sure about any of that... but you could view their current driver as a software OoOE-like implementation, since according to that slide their driver can choose to run an in-order command (e.g. anything in DX11 and earlier iirc) as out of order as the driver interprets the commands and determines it would not break the result guarantee.

 

[sze5003]

AMD will never achieve OC like Maxwell even Pascal cannot.

GTX 970 is still a better card and glad to know that. I fell sorry for those who sold their R9 390 or X for 480. it is downgrade in term of performance.

I wish people would stop making claims of better than this or better than that. At least until actual reviews are out and benchmarks from most games people play.

I don't care if it's better. I only care if it's better than what I currently have.

 

[lukart]

AMD will never achieve OC like Maxwell even Pascal cannot.

GTX 970 is still a better card and glad to know that. I fell sorry for those who sold their R9 390 or X for 480. it is downgrade in term of performance.

They probably thought they would get GTX980 performance

Btw guys, the 480 is 970 /390 performance on average, not even 390X.
In few games match the 390X and other few lower than 970.

There will be a few tricks to improve the performance even with out overclock (running stock clocks) , perhaps you guys could guess?

Stop day dreaming

 

[zinfamous]

AMD will never achieve OC like Maxwell even Pascal cannot.

GTX 970 is still a better card and glad to know that. I fell sorry for those who sold their R9 390 or X for 480. it is downgrade in term of performance.

well, you don't know any of this yet.

more likely than not, your assessment will be correct about games that are older and getting older every year, while 480 and other Polaris/Vega chips will continue to improve and outperform with, well, actual new games in DX12 compared to what nVidia has locked themselves into for the next 3 or so years.

Prepare to be disappointed with your obsolete purchase 4 months ago?

 

[desprado]

At least we will have a proper review with in few days. No down clocking of nvidia card like Roy wanted, no blocking of boost clocks, benchmark on variety of games.

 

[MTDEW]

For those that care. (im sure somebody does....LOL )
Doom RX 480 Gameplay FPS Test

 

[PhonakV30]

AMD will never achieve OC like Maxwell even Pascal cannot.

GTX 970 is still a better card and glad to know that. I fell sorry for those who sold their R9 390 or X for 480. it is downgrade in term of performance.

RX 480 is a better card and will consume ~110w and feel sorry for those who have GTX 970, this card will not have problem 3.5GB like GTX 970.once DX12 come out with Async compute , GTX 970 is no longer existed!

 

[William Gaatjes]

Yeah I wouldn't take those slides to mean OoOE. That's a somewhat specific thing but at the same time it can mean a lot of different implementations.

Key to the OoOE concept is that you still get back the same result as if you had done the operation in a truly blocking based serial manner, except (transparently) it is not actually truly serial.

Here, they are exposing a different feature. They are allowing you to forgo the guarantee of the same result from the hardware/driver as an in-order operation. As it explains, this is not a problem where you as the developer know that in-order or out-of-order results for your particular command will be identical. But if you ran a command that did need to be completed in order you might get race conditions or an incorrect result, whereas with OoOE you would not.

It's a related but distinct concept.

Missed your post. We are out of sync.
If you are referring to the out-of-order rasterization, i understand what they do.

It was more the rumor about that the the CU might have some more out of order execution logic in the vector unit and scalar unit as that might explain some rumors why it is so fast. It is probably from better utilization of all the CU's as commanded and controlled by the improved command processors.

 

[Mercennarius]

AMD will never achieve OC like Maxwell even Pascal cannot.

GTX 970 is still a better card and glad to know that. I fell sorry for those who sold their R9 390 or X for 480. it is downgrade in term of performance.

You can buy TWO brand new RX 480s for $399....the price of what one 390X sold for recently....and you end up with a crossfire setup that's FASTER than a 980 Ti and probably not far from the new 1080 all while using as much power as a single card does at that performance level...how does that not impress you?

 

[dzoni2k2]

At least we will have a proper review with in few days. No down clocking of nvidia card like Roy wanted, no blocking of boost clocks, benchmark on variety of games.

Pascal needs no Roy to downclock. It does it very well by itself.

 

[Shivansps]

Witcher 3 seems bad, becuase its not full even at max settings and it drops to 50 in a light areas... but is stupid to test AMD cards with hairworks on on W3!...

A GTX970 does about 40-45 in the same area at maximum settings.

 

[TeknoBug]

RX 480 already popping up on Canadian vendors for about $370-400, if it's indeed launching at around $199 USD then $370 CDN is a massive ripoff ($199 USD = $255ish CDN), I'd get one if it was $279-$329 CDN.

 

[boozzer]

For those that care. (im somebody does....LOL )
Doom RX 480 Gameplay FPS Test

I care 480 is gonna be the value king for 2 or more years it seems if everything holds.

RX 480 already popping up on Canadian vendors for about $370-400, if it's indeed launching at around $199 USD then $370 CDN is a massive ripoff ($199 USD = $255ish CDN), I'd get one if it was $279-$329 CDN.

buy it on ebay bro. if there are no supply problems, you will see alot of it on ebay for close to msrp, not 100$ more.

 

[desprado]

edited.

 

[xthetenth]

well, you don't know any of this yet.

He knows all of it, the question is whether he's backed by reality.

There's so much FUD circling, it's a total mess.

 

[Headfoot]

Missed your post. We are out of sync.
If you are referring to the out-of-order rasterization, i understand what they do.

It was more the rumor about that the the CU might have some more out of order execution logic in the vector unit and scalar unit as that might explain some rumors why it is so fast. It is probably from better utilization of all the CU's as commanded and controlled by the improved command processors.

Gotcha, it will definitely be interesting to see the technical deep dive on Polaris when the NDA is up

 

[dzoni2k2]

RX 480 already popping up on Canadian vendors for about $370-400, if it's indeed launching at around $199 USD then $370 CDN is a massive ripoff ($199 USD = $255ish CDN), I'd get one if it was $279-$329 CDN.

Its a socialist country, like most of Europe. Someone has to pay for the free healthcare and education.

 

[sze5003]

Witcher 3 seems bad, becuase its not full even at max settings and it drops to 50 in a light areas... but is stupid to test AMD cards with hairworks on on W3!...

A GTX970 does about 40-45 in the same area at maximum settings.

He did one more video without hair works. I forget what my card gets on Witcher 3. Everything is highest except for shadows and water, lights.