AMD Polaris architecture – GCN 4.0

[TheRealDeal]

Its not putting up a good fight here in ROTTR is it :-

http://hexus.net/tech/reviews/graphics/93050-nvidia-geforce-gtx-1070-founders-edition-16nm-pascal/?page=6

Getting spanked off the 980, only beats it at 4K by a single frame on avg!

At Tpu it's not half as bad in TOTTR. Still gets beaten at 1080p by the gtx980 but come 1440p it's under 10% slower than the ti and nearly on par at 4k.

https://www.techpowerup.com/reviews/NVIDIA/GeForce_GTX_1080/21.html

Overall it's holding up well against the stock 980ti

https://www.techpowerup.com/reviews/NVIDIA/GeForce_GTX_1080/26.html

 

[andybird123]

Hah I don't even know how it got there. Will remove it. Well, everyone is basically saying that Pascal is just as limited when it comes to dx12 as Maxwell, so we'll see once we get real dx12 games like bf1.

I'm not sure who "everyone" is, but from looking at the benchmarks available Pascal is getting a boost in DX12 from DX11, so I don't think its entirely fair to say that Pascall s "just as limited as maxwell". It is still the fastest DX12 card available right now, and it doesn't sounds like Polaris is aiming for taking the crown.

I guess it depends what AMD do with FuryX prices while we wait for Vega. 20% faster than a FuryX even in DX12 / AMD sponsored games is still a decent amount.

 

[Mauller]

Just becasue it is computex and AMD related.

BAM! http://www.fudzilla.com/news/processors/40784-amd-has-zen-ready-for-demo

 

[Toss3]

Just becasue it is computex and AMD related.

BAM! http://www.fudzilla.com/news/processors/40784-amd-has-zen-ready-for-demo

Holy ****! YES! This just got 1000 times more interesting! Cannot wait for tomorrow (tonight)!

 

[Sixense]

It funny back in 2015 AMD boasted about tiny card and now AMD going to show off absolutely massive card much longer than Radeon Pro Duo's 270mm length lol.

360mm or 14.17 inch card certainly wont fit in most cases included mine.

3x120mm rad for an AIO cooler. Judging by the size of the block I expect the card itself to be a bit longer than the Nano, something like 9-10 inch long maybe?

Rad got me interesting, if P10 is like 130W card, and general idea for wc I've seen is a 120mm for every 100W, 3x120 mean something like 250+W consumption on OC?

 

[Seanspeed]

Hah I don't even know how it got there. Will remove it. Well, everyone is basically saying that Pascal is just as limited when it comes to dx12 as Maxwell, so we'll see once we get real dx12 games like bf1.

"Everyone is saying"? No, just a few AMD fans who obviously haven't looked at the benchmarks and just want to repeat misinformation because it suits their agenda.

Again, not saying it's up to par with AMD, but DX12 performance with Pascal has definitely improved over Maxwell. It is undeniable.

 

[Sacrifice]

Just a quick note to everyone trying to ascertain something from the rad size. The article specifically states that the size used in this showcase shouldn't be used to predict anything about polaris. If you had a product showcase, would you pick the largest rad to show how capable the cooling solution is, or the smaller rad?

 

[r22snapper]

Thanks for all the Pascal talk, It'd be a real pest having to go and read that thread as well

 

[bayo000]

http://www.techpowerup.com/222988/powercolor-next-gen-lcs-liquid-cooling-solutions-pictured

360x120mm cooler for single GPU. They must be overclockers dreams for it to be that big

Or it just needs so much cooling to run at stock

 

[ICDP]

"Everyone is saying"? No, just a few AMD fans who obviously haven't looked at the benchmarks and just want to repeat misinformation because it suits their agenda.

Again, not saying it's up to par with AMD, but DX12 performance with Pascal has definitely improved over Maxwell. It is undeniable.

I don't see anyone claiming the DX12 performance is not better than Maxwell, of course it is. The issue we are seeing is that in DX12 the gap remains static between 980Ti and 1080, which point to the fact that there are no improvements on how DX12 is handled in architecture/design with Pascal.

Wiki is now showing that neither Maxwell nor Pascal have Asynch Compute hardware (changed from asynch shaders in older table to avoid confusion). To me that indicates that Pascal is a die shrunk Maxwell.

https://en.wikipedia.org/wiki/Feature_levels_in_Direct3D

Scroll down to the DX12 Feature Level table. It is clear that Pascal has a lower DX12 feature set than Fiji and while we still have no idea of Polaris DX12 feature levels it's a safe bet it will be at least as good as Fiji.

This may not be a big problem now but the problem is that DX11 will become more and more irrelevant before this year is out for AAA titles. Failure to recognise or accept this potential future proofing issue is not a reflection on those pointing out that Pascal seems backward facing (ie DX11). It's a reflection on those intent on ignoring , or dismissing it as irrelevant.

Note: I am very impressed with 1070 and the planned price point, far more so than 1080. 1070 brings 980Ti/Fury X performance for a singinficant drop in price but the fact it's still behind Fiji in most DX12 titles is a concern IMHO. I have no idea if Polaris will improve upon Fiji for DX12 and I'm not saying AMD > Nvidia overall. I'm pointing out that IMHO (from evidence) that Pascal is a die shrunk Maxwell and as such could end up a poor(er) choice for future DX12 titles. I'm playing the waiting game and this could be wrong but the indications are AMD > Nvidia for DX12.

 

[Gregster]

Not a particularly good example when the game clearly appears to favour Nvidia cards or has gimped AMD. The 780ti is matching the FuryX which should not be the case at all if the code was good. While AMD sponsored games don't have such an obvious discrepency, Nvidia games usually do which is sadly a reflection of how the company operates.

I am sure I have heard that before somewhere

 

[Seanspeed]

I don't see anyone claiming the DX12 performance is not better than Maxwell, of course it is.

That is *exactly* what was claimed.

The issue we are seeing is that in DX12 the gap remains static between 980Ti and 1080

You are not 'seeing' that anywhere, though. That's what I'm trying to say. The gap IS actually bigger between the two in DX12 titles than in DX11 titles. Hence - Pascal has improved DX12 performance.

http://i.imgur.com/FQKfSxQ.png

http://i.imgur.com/Qi5qJ0N.png

Check any of the reviews. They all corroborate this. You obviously haven't paid close enough attention.

This may not be a big problem now but the problem is that DX11 will become more and more irrelevant before this year is out for AAA titles.

This just wont be the case. We'll see a few more DX12 implementations, but DX11 is hardly on any dramatic collision course with irrelevancy. It'll stick around for quite a while. Especially with DX11.3 having quite a lot of the features introduced with DX12.

Either way, you're trying to make this into some AMD vs Nvidia ****ing match even after I've tried very hard to make it known in my posts that it wasn't my intention. Only to rubbish the notion that Pascal is 'just as' hamstrung in DX12 as Maxwell was.

 

[Ian Evey]

I am sure I have heard that before somewhere

Think i have seen the same thing here > http://www.guru3d.com/articles_pages/hitman_2016_pc_graphics_performance_benchmark_review,6.html

Is this a sponsored amd game?

 

[h4rm0ny]

Thanks for all the Pascal talk, It'd be a real pest having to go and read that thread as well

Yes! PLEASE can we reduce the amount of Nvidia / Pascal crap in this thread. Not everything has to be a football match!

 

[Toss3]

That is *exactly* what was claimed.


You are not 'seeing' that anywhere, though. That's what I'm trying to say. The gap IS actually bigger between the two in DX12 titles than in DX11 titles. Hence - Pascal has improved DX12 performance.

http://i.imgur.com/FQKfSxQ.png

http://i.imgur.com/Qi5qJ0N.png

Check any of the reviews. They all corroborate this. You obviously haven't paid close enough attention.


This just wont be the case. We'll see a few more DX12 implementations, but DX11 is hardly on any dramatic collision course with irrelevancy. It'll stick around for quite a while. Especially with DX11.3 having quite a lot of the features introduced with DX12.

Either way, you're trying to make this into some AMD vs Nvidia ****ing match even after I've tried very hard to make it known in my posts that it wasn't my intention. Only to rubbish the notion that Pascal is 'just as' hamstrung in DX12 as Maxwell was.

You couldn't have chosen a worse "dx12" title to prove the improvements. Tomb Raider doesn't support async compute, nor is it a purely DX12 title. The reason being that it is an nvidia title, so coincidentally async was dropped.

Here you can see more results:
http://www.tomshardware.com/reviews/nvidia-geforce-gtx-1070-8gb-pascal-performance,4585-4.html

As you can see the Fury X edges out the GTX 1070 in AotS and Hitman, and comes even dangerously close to the GTX 1080. The added hardware scheduler is definitely helping here, as pascal cards are seeing a performance improvement in dx12 over dx11, while maxwell cards usually see the performance go down (could be software differences as well). So yes, pascal cards do seem to benefit more from dx12 than maxwell, but AMD are still the only ones fully supporting async and that is something that is going to give Vega and Polaris an edge.

Oh, and I'm just hoping AMD will have the advantage as I really want to see prices go down and AMD being competitive once again! And I really see no harm in comparing Nvidia and AMD considering they are the only two players on the market currently, as long as it is in good fun.

 

[Gregster]

Think i have seen the same thing here > http://www.guru3d.com/articles_pages/hitman_2016_pc_graphics_performance_benchmark_review,6.html

Is this a sponsored amd game?

Yes and oddly enough, AoTS as well. Both AMD sponsored but I give credit to AMD for working closely with the devs and don't feel the need to sling mud. Good job AMD and in those two DX12 games, AMD performance is top notch

 

[D.P.]

Most of the DX12 performance increase comes from sheer clockspeed increase rather than any architectural improvement. Most reviews are comparing a 1080 @ 1800MHz boost vs a 980ti @ 1300MHz boost which accounts for the difference. Run the 1080 at the same speed as an overclocked 980ti (e.g 1600MHz) and they will be pretty much similar in a DX12 game like Hitman.

If you drive a Nissan micra and Ferrari Enzo at 60mph they both travel at the same speed. Does this mean a Nissan micra is actually a super car?


All his talk of DX12 is really meaningless. It's just an API to access the hardware. DX11 had limitations but nvidia proved how you can overcome those limitations. The fact that DX12 does gain nvidia card much performance is not a problem of nvidia cards, it is a problem of AMD drivers and GPUs. Something which Polaris is supposedly helping to fix. Therefore Polaris GPUs will likely have a smaller difference between DX11 and 12, does this mean Polaris is a step backwards?

IMO as a developer AMD spending time on game-dev forums, you won't see a big jump in DX12 uptake soon. The biggest AAA game engines are slowly moving that way, games based on those engines will take time. Performances differences are going to be small unless you have a severely under performing CPU.

 

[D.P.]

I believe reason is game isn't using any dx12 features that benefit amd.

You mean it isn't biased then?

 

[Seanspeed]

You couldn't have chosen a worse "dx12" title to prove the improvements. Tomb Raider doesn't support async compute, nor is it a purely DX12 title.

Since when is async compute the only thing that defines whether something is DX12 or not? You've obviously been reading tons of propaganda on this.

Either way, I posted a Hitman bench, too. One you conveniently decided to ignore in your attempt to rubbish my example.

As you can see the Fury X edges out the GTX 1070 in AotS and Hitman, and comes even dangerously close to the GTX 1080.

Completely irrelevant to everything I was talking about.

So yes, pascal cards do seem to benefit more from dx12 than maxwell

All I was trying to say. Could have saved us all a lot of time by just agreeing with me in the first place instead of this whole song and dance.

 

[Phixsator]

If you drive a Nissan micra and Ferrari Enzo at 60mph they both travel at the same speed. Does this mean a Nissan micra is actually a super car?

You are missing the point. Running both cards clock for clock would highlight any architecture differences between the two which would include how well Pascal actually handles DX12 compared to maxwell. Now of course the results wont be 100% accurate but it would give a better estimate.