AMD Polaris architecture – GCN 4.0

[SiDeards73]

hints coming out of the press event are pointing towards exceptional performance for the gpu.

Got any links? Interested to read more

 

[AllBodies]

hints coming out of the press event are pointing towards exceptional performance for the gpu.

IMO the most interesting thing will be how the performance per watt stacks up against Pascal, rather than necessarily the absolute performance difference.

Reason being, it gives an indication from a tech perspective of whether 14LPP or 16FF+ is better, and then that gives bearing on what both Vega and Zen can be like.

If Polaris has exceptional performance/W, then Vega should be nuts. Since it'll have HBM2 cutting it's power usage, along with possibly being a further revision of GCN.

 

[TheRealDeal]

Not sure how its out of place, I think you're just offended by it. I want the best card for £350 and I'm hoping AMD provide that with Polaris, if they dont and the 1070 is the best card for £350 then I'm going to be kicking myself for investing in Freesync

Its perfectly in line with the thread as I and many others are in the same boat whilst waiting for Polaris news

The problem you have is a slightly slower Polaris would still give you an overall better experience than a slightly faster 1070 due to Freesync.

 

[Jono8]

It's true, you seem a bit butthurt anytime anyone suggests anything anti Nvidia in this and other threads, it's just an observation.

A bad observation. I just like to correct people when they are wrong. There is no basis in precedent or fact that even hints that the 1070 wont clock as well as the 1080.

I literally couldnt care less what brand of card i have in my pc. I will buy the cheapest card in the performance bracket i want, that is available at the time

 

[SiDeards73]

I'm not convinced by HBM given how bad the Fury cards were in regards to over clocking and the fact they only really excelled at higher resolutions and we're pretty bad at lower resolutions.

If they can leverage better 1080p performance and then have it scale up further to 4k and beyond with HBM I will be sold, the largest part of the market is still 1080p and saying Fury was meant for the high end only is not smart, people still expect it to perform better the lower the resolution not the opposite lol.

 

[fs123]

Got any links? Interested to read more

Yes we need links to these hints. The Hitman demo indicated good performance so surely AMD couldn't have gone backwards since then.

 

[Relentless81]

The problem you have is a slightly slower Polaris would still give you an overall better experience than a slightly faster 1070 due to Freesync.

Interesting point, it would still need to be a worthy upgrade from a 290X but yes you're right if its faster than a 290X and slower than a 1070 for the right price it could still be a worthwhile purchase

 

[TheRealDeal]

I'm not convinced by HBM given how bad the Fury cards were in regards to over clocking and the fact they only really excelled at higher resolutions and we're pretty bad at lower resolutions.

If they can leverage better 1080p performance and then have it scale up further to 4k and beyond with HBM I will be sold, the largest part of the market is still 1080p and saying Fury was meant for the high end only is not smart, people still expect it to perform better the lower the resolution not the opposite lol.

Dx12 seems to let the Fury X stretch it's legs at 1080p. I think the problems at 1080p was more down to driver overhead in dx11 than anything HBM. On the overclocking we will just have to see if it was the Fury chip or the HBM design.

 

[Mauller]

I doubt it.

GCN (even though it's a new version) clocks lower than Kepler and Maxwell on 28nm. And 14nm is meant to be slightly denser and more power efficient than 16nm, but not clock as high.

So AMD are likely going for power efficiency and IPC rather than clockspeed.

The 14nm samsung process is rated for a 2ghz max clock at its max density, 28nm from tsmc is rated at 1-1.2 GHz at max density. The reason nvidia have gotten away with higher clocks is because they never go near the max density of the process.

 

[SiDeards73]

I bet quite a few of waiting on upgrades are still running 290 as AMD has not really had anything else yet that matches it in the price / performance area, I know the only reason I have not upgraded yet is due to this and the fact my 290 still runs superb in D3, Division, WoW and ESO which are the only games I really play

If the 480x hits the right ratio I will buy one even if it's a side or marginal upgrade until Vega, much like 980ti owners are doing with 1080. Mostly because the new GCN will offer better benefits going forward also with dx12

 

[D.P.]

He's assuming a cut down 1080 die is effectively 75% of the full size putting it around same size. This completely ignores the benefits of having dead silicon and the effect on yields so its a bit of a wonky assumption.

And completely ignores the fact that not 25% of the entire chip is disabled.

 

[AllBodies]

The 14nm samsung process is rated for a 2ghz max clock at its max density, 28nm from tsmc is rated at 1-1.2 GHz at max density. The reason nvidia have gotten away with higher clocks is because they never go near the max density of the process.

Well hopefully then.

Obviously I want to see AMD do as well as possible since it's good for everyone. I was just speculating on previous cards.

 

[D.P.]

Polaris 10 is made on a denser process so massaging of figures does not change the fact that Nvidia is using a 314mm2 chip and Polaris 10 is a 232mm2 one.

We have some new leaks about Polaris can we at least talk about that?

When I did some research on this a few days ago it indicated TSMC16nm was denser than Samsung 14nm. Happy to be proven wrong. Yet no where can I find anything to collaborate that samsung/GF is actually denser on the LPP process.

 

[Jono8]

I bet quite a few of waiting on upgrades are still running 290 as AMD has not really had anything else yet that matches it in the price / performance area, I know the only reason I have not upgraded yet is due to this and the fact my 290 still runs superb in D3, Division, WoW and ESO which are the only games I really play

If the 480x hits the right ratio I will buy one even if it's a side or marginal upgrade until Vega, much like 980ti owners are doing with 1080. Mostly because the new GCN will offer better benefits going forward also with dx12

Indeed. With the 1080 being stupid money i will get a 1070 or the top polaris card depending on which is faster and what the price differential is . Will be a nice stop gap until the big boys arrive.

 

[AllBodies]

And completely ignores the fact that not 25% of the entire chip is disabled.

I'm not trying to suggest the top P10 will be significantly faster or anything.

I'm just pointing out that if you take the net different of the 1070 being cut-down, and 14nm being denser, clearly they're going to be very close together (on paper anyway).

 

[D.P.]

Why? It will likely clock just as well as the 1080 (so 2ghz+ easy)

It is likely to clock significantly less because Pascal has been engineered to clock well at the expense of IPC. AMD have increased IPC which will undoubtedly affect max clocks because the 2 are intrinsically linked.

 

[Jono8]

It is likely to clock significantly less because Pascal has been engineered to clock well at the expense of IPC. AMD have increased IPC which will undoubtedly affect max clocks because the 2 are intrinsically linked.

Argh!

Read the rest of this page.

 

[D.P.]

I'm not trying to suggest the top P10 will be significantly faster or anything.

I'm just pointing out that if you take the net different of the 1070 being cut-down, and 14nm being denser, clearly they're going to be very close together (on paper anyway).

But you are ignoring that way less than 25% of the 1070 is disabled. 25% of the CUs are disabled but that is a much smaller,percentage of the entire GPU. A lot of die area exist in things like video encoding, ROPs, memory interface, GPU front-end that exist in the entirety.

 

[MyBrainz]

Maybe this is one of the 'hints' about Polaris..

http://www.pcworld.com/article/3076...ases-radeon-480-graphics-card-on-twitter.html

Sorry if already posted

 

[Mauller]

It is likely to clock significantly less because Pascal has been engineered to clock well at the expense of IPC. AMD have increased IPC which will undoubtedly affect max clocks because the 2 are intrinsically linked.

They have increased core utilisation more than the ipc of the core itself from the features they have added. The density the design is made at also plays a major part in clock scaling.