AMD Polaris architecture – GCN 4.0

[pmc25]

What are next weeks lottery numbers, thanks!

Trolling as usual, I see.

You've been around long enough to know that will be the case, and why.

For others .. aside from large numbers being sold as add-in boards, huge numbers will go to OEMs (including vast quantities to Apple). PS4.5 will have a P10 too, and NX / XB2 may have derivatives.

There hasn't been another chip in recent memory that has remotely near the potential production numbers of Polaris 10. GP104 has no such prospects. The only one that comes close is the P11.

 

[Rroff]

Now, the head of the AMD Radeon Group, Raja Koduri, has actually admitted this problem with thier GPU's, and has said they have completely rebuilt the front end of their GPU's for Polaris to fix this problem.

It is kind of funny that with Polaris/Pascal nVidia has gone a little more towards an AMD style and AMD has made some changes that are a little more nVidia style.

 

[queamin]

It is kind of funny that with Polaris/Pascal nVidia has gone a little more towards an AMD style and AMD has made some changes that are a little more nVidia style.

Yes but Amd need to do that as it was/is hard to get the most out or their cards the max flops on a 390/x is about the same as 980ti but 980ti beats it easy quite a lot of the time so Amd need to change Polaris a lot.

 

[fs123]

Even the 290/390 has a DX11 CPU batching problem, i know that for a fact as i went from a 290 to a GTX 970 and given that i develop my own games i can see whats going on in the background.

The truth of the matter is Maxwell is far more efficient at batching, the result is Maxwell can push about 70% more drawcalls in DX11 than AMD can.

The reason why its far more pronounced with Fiji than it is with Hawaii-XT at 1080P is because Fiji is a massively more powerful architecture.
This is why you often see a Fiji GPU only a little faster than a Hawaii-XT, at 1080P, at higher resolution like 4K where the power of the GPU is more important than Drawcalls Fiji comes into its own.

As a GPU its actually more powerful than a GTX 980TI, Fiji's problem is the 980TI is able to use more of its rendering power than Fiji can, at least at 1080P in DX11.

Now, the head of the AMD Radeon Group, Raja Koduri, has actually admitted this problem with thier GPU's, and has said they have completely rebuilt the front end of their GPU's for Polaris to fix this problem.

This is the interesting aspect of it all. If AMD cards can keep up with Nvidia Maxwell cards even with the huge 70% overhead then what will Polaris achieve with that same bottleneck gone? Theoretically a Fiji without the overhead should be more powerful than a TitanX and in some of the recent DX12 games like Hitman we can see that as proof.

 

[Orangey]

VR is priced out of range for 99% of people, you can't even buy them yet and they don't have any games.

If you're hoping AMD will outsell NV based on VR hype you're wrong. How many people are going to buy a bargain bin GPU to run a £500 headset? Before the price was announced casuals were snapping up 970s to be "VR ready", are they going to sell up and buy a glorified 390X instead to be "VR readier" for another 6 months in the vague hope the headsets might get a massive price drop?

 

[Mauller]

It is kind of funny that with Polaris/Pascal nVidia has gone a little more towards an AMD style and AMD has made some changes that are a little more nVidia style.

Not really on the AMD side, its just a fix in the scheduler to improve single threaded throughput. It is still highly parallel as before. One of the big improvements to the cores I instruction prefetching.

And we have no idea if Pascal has AMD levels of async capability. They would have bigger it up at the release event.

 

[LoadsaMoney]

Loads if you look at the 390 and what it can do in dx12 on 28nm what do you think polaris will do on 14nm with all it improvements if they mess this up they deserve to go bankrupt as dx12 and 16/14nm is what Amd have been waiting for for their cpu's and gpu's.

We havn't seen it do anything in Dx12 yet, as all we've got is a couple of games that have had Dx12 patches added to them, we can't possibly make a judgement on that imo, when we get proper actual Dx12 games, that havn't just been patched later, then we'll see.

 

[Mauller]

We havn't seen it do anything in Dx12 yet, as all we've got is a couple of games that have had Dx12 patches added to them, we can't possibly make a judgement on that imo, when we get proper actual Dx12 games, that havn't just been patched later, then we'll see.

Doom will be getting a vulkan patch a few days after release.

 

[humbug]

It is kind of funny that with Polaris/Pascal nVidia has gone a little more towards an AMD style and AMD has made some changes that are a little more nVidia style.

Yes. good observation...

AMD never saw the benefits of batching, as they explain in this video, watch it all but the crucial bit is at about 1:30m.

Instead they went for A-Synchronous Compute, they explain why in the video.

Nvidia did see the benefit, for DX11 at least, AMD were looking to Mantle and DX12.

Now that Nvidia have done it AMD are looking at it again.

 

[eazyrider]

Will polaris have hbm?

GDDR5, Vega will have HBM2

 

[AlamoX]

VR is priced out of range for 99% of people, you can't even buy them yet and they don't have any games.

If you're hoping AMD will outsell NV based on VR hype you're wrong. How many people are going to buy a bargain bin GPU to run a £500 headset? Before the price was announced casuals were snapping up 970s to be "VR ready", are they going to sell up and buy a glorified 390X instead to be "VR readier" for another 6 months in the vague hope the headsets might get a massive price drop?

thats the point, the argument of palmer(rift) that it's useless to sell headset cheaper, is that most of the gaming rigs are way lower than the required specs, and thats from the guy making the headset, usualy if the required spec average spread, the headset price will go down, so does the cost overtime, but for now they want to milk ppl first.
there is a reason why oculus and HTC doesnt mass produce the headset, man they do pre-orders for like 10-15k units per wave, that should tell you we are nowhere near cost efficient production of the headset.

 

[queamin]

We havn't seen it do anything in Dx12 yet, as all we've got is a couple of games that have had Dx12 patches added to them, we can't possibly make a judgement on that imo, when we get proper actual Dx12 games, that havn't just been patched later, then we'll see.

Yes and in a earlier post on another thread I said the same but to be honest at moment with consoles it is looking as if Amd is prepared for dx12 more than Nvidia now what ever comes after pascal will be interesting as it will be build from the ground up for dx12 I think from both camps.

 

[pmc25]

GDDR5, Vega will have HBM2

AMD have said that Polaris' memory controller is compatible with GDDR5(X), HBM and HBM2. They said HBM2 wouldn't appear until Q1 next year due to price and availability (with Vega), but they never said there wouldn't be an HBM Polaris. I'm sure they've designed them and done test chips ... but if Pascal is as weak as it looks, I suspect they'll skip it and / or a medium sized Polaris chip entirely (Polaris 10 is quite a bit smaller than GP104).

 

[humbug]

This is the interesting aspect of it all. If AMD cards can keep up with Nvidia Maxwell cards even with the huge 70% overhead then what will Polaris achieve with that same bottleneck gone? Theoretically a Fiji without the overhead should be more powerful than a TitanX and in some of the recent DX12 games like Hitman we can see that as proof.

Both Fiji and Hawaii-XT are very powerful architectures, more powerful than it appears in most cases.

 

[humbug]

Both Fiji and Hawaii-XT are very powerful architectures, more powerful than it appears in most cases.

When not let down by idle shaders suffocated by saturated call instruction queues.

 

[Panos]

Not really on the AMD side, its just a fix in the scheduler to improve single threaded throughput. It is still highly parallel as before. One of the big improvements to the cores I instruction prefetching.

And we have no idea if Pascal has AMD levels of async capability. They would have bigger it up at the release event.

I tried to find that info also. The only mentioned it once for "driver" level, nothing was about hardware level. And that is a very very big difference.

Is like saying that they can make a 4 cylinder engine to sound like a V12 by just fiddling with the exhaust.

 

[Panos]

When not let down by idle shaders suffocated by saturated call instruction queues.

For those might consider where the 290X/290 stand on that chart, replace the 390X/390 if you have overclocked your 290s to the same speeds as the stock 390s...
Because they are the same card

 

[CAT-THE-FIFTH]

It is kind of funny that with Polaris/Pascal nVidia has gone a little more towards an AMD style and AMD has made some changes that are a little more nVidia style.

I kind of mentioned this last year too - I can see both converging in a bit in the way they do things.

 

[humbug]

For those might consider where the 290X/290 stand on that chart, replace the 390X/390 if you have overclocked your 290s to the same speeds as the stock 390s...
Because they are the same card

Well, lets put it this way, Hawaii has not become any more powerful from renaming it 390/390X, not even so much from its 5% overclock.

Its moved up the chart even in DX11 (now competing with a GTX 980) because AMD have spent the last 2 years tweaking thier driver overheads 'as much as they can in software' as to allow more of Hawaii's actual potential to be unlocked.
Its still a long way behind Nvidia solution tho, make no mistake, a long way.

 

[queamin]

People who still have a single 290/x have done well as it looks good for dx12.