AMD Polaris architecture – GCN 4.0

[CAT-THE-FIFTH]

Anandtech is a horrible example to use - the bickering was much worse than here and they had to split the GPU forums into AMD and Nvidia parts.

 

[bru]

Serious question, do you think that HBM2 cards will suffer the same shortfalls as the Fury cards did with 1080p performance in relation to 1440p and 4k ?

Obviously this would probably effect both AMD and NVidia. Could this mean that the next gen are not going to be optimal for 1080p users ? (myself I'm on 1200p so pretty much the same)

 

[Mauller]

Serious question, do you think that HBM2 cards will suffer the same shortfalls as the Fury cards did with 1080p performance in relation to 1440p and 4k ?

That was nothing to do with HBM, more to do with driver overhead.

 

[humbug]

Serious question, do you think that HBM2 cards will suffer the same shortfalls as the Fury cards did with 1080p performance in relation to 1440p and 4k ?

Obviously this would probably effect both AMD and NVidia. Could this mean that the next gen are not going to be optimal for 1080p users ? (myself I'm on 1200p so pretty much the same)

That was nothing to do with HBM, more to do with driver overhead.

This ^^^^ and i hope AMD are fixing that, its not the performance of Fury-X that's the problem, it beats the 980TI consistently at 4K where DrawCall overheads are not an issue.

DX12 will sort that completely but they cannot depend on it as its going to be a while before DX12 is the every game standard like DX11 is now.

So they have to do whatever it take to get at least DrawCall parity with Nvidia.

 

[tommybhoy]

+1, don't sort it, suffer the consequences AMD.

Imo, if big Polaris doesn't hit the ground running with anything but stellar release day driver optimisation, then they have learned nothing and same old cycle continues.

 

[redrooster303]

Same argument still on going? Think I'll come back later

 

[Orangey]

I just had a great idea.

 

[~Mike~]

Then don't, 99% of anything interesting on this forum gets posted to a new thread. If banter annoys you just stick to reading opening posts.

I don't, I look elsewhere for info these days, some of it here is banter but most of it is rubbish and no I don't get annoyed.

You seem to post that quite a bit in this sub section.

oh Greg you make me laugh, you might want to get your facts straight.

probably because its true? We've had 10 pages of bitter arguments between people on one side of the "fence", or the other... plus we've had CPU arguments too. 90% of this thread is nothing to do with Polaris.

Exactly.

 

[FoxEye]

See here is where you are putting your own agenda and assumptions ahead of the actual facts. Your preconceived notion of "big Polaris" is just that, a notion not based on any facts or even rumours. Should AMD be releasing a Polaris GPU with a die size similar to Fury X in 2016 for you to be happy? Only once a process node matures will we start seeing such large die sizes again. The first year on 16/14nm was never going to see the large die GPUs released.

It's obvious what "big Polaris" means, isn't it? The one with the most transistors they can squeeze into a given process. That means FuryX from this gen; it means Titan X from nV.

We know that "tiny Polaris" and middle Polaris are coming in 2016.

What is the argument about, again? You've just said that

The first year on 16/14nm was never going to see the large die GPUs released.

So why do you keep insisting that "big Polaris" could come in 2016? Either you've got your wires really, really crossed, or you've just contradicted yourself?

This isn't about your speculation it's about you making up "facts" so you can shoot them down to show how right you are. Nowhere did it say that the high end Polaris due in 2016 would have a similar or only slightly better performance than Fury X. Nowhere has any of these articles claimed "big Polaris has been delayed until 2017". Yet you made those erroneous statements then proudly proclaimed Polaris a failure because you read somewhere it's only as fast as Fury X.

This is the kind of "factual" statements you have made with zero shred of proof, just thoughts in your head you spout as facts.

If we're talking "facts", perhaps you ought to be more sure of yours. I didn't use the word "delayed", nor did I ever call Polaris a "failure". Or even imply this. And whilst I used the word "speculation" a lot, there's only a couple things I've presented as fact. The "two new GPUs in 2016" is definitely a fact - it comes from AMD themselves. "Console class" is likewise a fact.

The rest of it is all speculation.

You seem to be having a hard time understanding this.

 

[drunkenmaster]

+1, don't sort it, suffer the consequences AMD.

Imo, if big Polaris doesn't hit the ground running with anything but stellar release day driver optimisation, then they have learned nothing and same old cycle continues.

Nonsense, firstly software/algorithm writing simply doesn't work like that. Someone writing drivers on day 100 of working with Polaris will have more knowledge and experience than on day 50. Drivers get better over time, that is how life works.

Second, people keep forgetting one basic little fact, Nvidia 'cheated' with driver optimisation. Aside from the fact that they had massively more problems, more crashing and more joke TWIMTBP releases than ever in the past two years, they moved hardware scheduling off the gpu for Maxwell... these two things are not coincidence along with removing driver overhead.

They could just use more of the CPU, they were working around DX11 massively to improve driver overhead and gain a lot in draw call performance.... but they did this not on the GPU, but the CPU. However working so heavily to bypass DX11 problems, they also appeared to introduce more driver problems than ever which shouldn't be a surprise.

So AMD had worse driver overhead because they kept hardware scheduling on die(with Nvidia almost certainly going back to that with Pascal), but they also had, in my experience, basically flawless day one performance on a single GPU in every single game I played in the past two years.

It's amazing what people determine are great drivers and make grand proclamations about who has to improve what.

Guess what, my AMD experience of day one gaming in the past 2 years has been perfect. By and large from the complaints I see(excluding both SLI and xfire, which has gone massively downhill as DX11 wraps up) AMD have had far far fewer problems in the past two years. If any grand proclamations were to be made, maybe they should be about Nvidia addressing their drivers to deliver day one stability to their users for a change because their stability is unacceptable even if the performance is good.


The reality is bringing the hardware scheduling back on die will give them limited power to just use extra CPU power to get around DX11 problems. It's quite possible bringing it back onto the GPU will reduce their DX11 performance when the issue is driver overhead.

 

[andybird123]

Wow, the day that DM basically accuses Tommy of being an Nvidia fanboi

Have I actually just entered the twilight zone?

 

[meldarthx]

Wow, the day that DM basically accuses Tommy of being an Nvidia fanboi

Have I actually just entered the twilight zone?

most likely because he didn't but I'd like what you are smoking

 

[CAT-THE-FIFTH]

most likely because he didn't but I'd like what you are smoking

The glare from Polaris was too bright.

 

[Mauller]

The glare from Polaris was too bright.

At least polaris is visible. Pascal is turning up like dark matter at the moment, a lot of theory but no direct observations.

 

[Overlag]

We know that "tiny Polaris" and middle Polaris are coming in 2016.

Even a 4096 shader processor core would be smaller on a 14nm chip so could be classed "middle". Couple that with the optimisations AMD are doing plus a mhz increase and its still going to be faster?


Anyway who says we are going to go back to huge cores again, they were only really around because 28nm had to make do longer than it should have.

 

[flopper]

Nonsense, firstly software/algorithm writing simply doesn't work like that. Someone writing drivers on day 100 of working with Polaris will have more knowledge and experience than on day 50. Drivers get better over time, that is how life works.

Second, people keep forgetting one basic little fact, Nvidia 'cheated' with driver optimisation. Aside from the fact that they had massively more problems, more crashing and more joke TWIMTBP releases than ever in the past two years, they moved hardware scheduling off the gpu for Maxwell... these two things are not coincidence along with removing driver overhead.

They could just use more of the CPU, they were working around DX11 massively to improve driver overhead and gain a lot in draw call performance.... but they did this not on the GPU, but the CPU. However working so heavily to bypass DX11 problems, they also appeared to introduce more driver problems than ever which shouldn't be a surprise.

So AMD had worse driver overhead because they kept hardware scheduling on die(with Nvidia almost certainly going back to that with Pascal), but they also had, in my experience, basically flawless day one performance on a single GPU in every single game I played in the past two years.

It's amazing what people determine are great drivers and make grand proclamations about who has to improve what.

Guess what, my AMD experience of day one gaming in the past 2 years has been perfect. By and large from the complaints I see(excluding both SLI and xfire, which has gone massively downhill as DX11 wraps up) AMD have had far far fewer problems in the past two years. If any grand proclamations were to be made, maybe they should be about Nvidia addressing their drivers to deliver day one stability to their users for a change because their stability is unacceptable even if the performance is good.


The reality is bringing the hardware scheduling back on die will give them limited power to just use extra CPU power to get around DX11 problems. It's quite possible bringing it back onto the GPU will reduce their DX11 performance when the issue is driver overhead.

Yes AMD is superior in tech.
However their marketing team sucks.

Lucky enough Polaris is a brighter future.
I am excited for the new 14nm tech and the HDR screens as this year we have some new fun things to game with

 

[CAT-THE-FIFTH]

At least polaris is visible. Pascal is turning up like dark matter at the moment, a lot of theory but no direct observations.

 

[FoxEye]

Really wish we knew a little more about these cards. I was hoping that £300 would buy 980/ti performance, but I'm starting to think that's optimistic.

I'm just not sure if both AMD and nV at this point see the reduced power as a feature people will pay /more/ for, over the current generation.

So I'm kind of bracing myself for "bigish Polaris" to come in at £500 or so.

(for ICDP: this is not presented as fact, to my knowledge AMD have not spoken about pricing at all.)

 

[Sacrifice]

Really wish we knew a little more about these cards. I was hoping that £300 would buy 980/ti performance, but I'm starting to think that's optimistic.

I'm just not sure if both AMD and nV at this point see the reduced power as a feature people will pay /more/ for, over the current generation.

So I'm kind of bracing myself for "bigish Polaris" to come in at £500 or so.

(for ICDP: this is not presented as fact, to my knowledge AMD have not spoken about pricing at all.)

Isn't the point about power reductions that if you can halve the perf per watt, you can double the performance at the same consumption as current gen?

 

[ICDP]

Really wish we knew a little more about these cards. I was hoping that £300 would buy 980/ti performance, but I'm starting to think that's optimistic.

I'm just not sure if both AMD and nV at this point see the reduced power as a feature people will pay /more/ for, over the current generation.

So I'm kind of bracing myself for "bigish Polaris" to come in at £500 or so.

(for ICDP: this is not presented as fact, to my knowledge AMD have not spoken about pricing at all.)

I would suspect ~£320-£350 Polaris should be very close to stock 980Ti IMHO. This would make it a direct replacement in price for the existing 390X and fall right in line with ~30% performance increase across the price range.

IMHO £550 will get you high-end Polaris and I expect ~30% faster than Fury X/980Ti at 4K.

These are the numbers (give or take 5%) I am expecting and anything more is a bonus.