AMD Polaris architecture – GCN 4.0

[Dave2150]

so why do people think AMD's stuff is coming first??
wasnt so long ago people were saying nvidia's is 1st maybe as early as feb/march?

im surprised we seeing tech demos so early but i dont think we should read too much into that

Remember that AMD have exclusive priority access to HBM2, due to them helping develop the technology.

NVIDIA will only have sloppy seconds, hence may choose to adopt a more relaxed release schedule.

 

[Steampunk]

And will this have full DX12 support?

I can't see how it wouldn't if it's arriving a year after DX12.

 

[GoogalyMoogaly]

No but you did say this.



The part in red is the salient point here. It was most certainly NOT "everyone" who was saying "nobody cares about power consumption". A few people may have but not "everyone". Now do you see why "some of us" but not "everyone" took the time to respond to your ludicrous claim?

You see the bit where I put it in quotes to imply that I didn't actually mean everyone?

 

[andybird123]

Remember that AMD have exclusive priority access to HBM2, due to them helping develop the technology.

NVIDIA will only have sloppy seconds, hence may choose to adopt a more relaxed release schedule.

Except there's also more than one supplier for HBM now and AMD only have priority at one of them.

 

[Steampunk]

How useful is HBM without the interposer? Who owns that?

 

[pmc25]

Thats assuming they are using HBM2 for the small chip and not using GDDR5/GDDR5X which is cheaper and higher volume.

The GF/Samsung process is meant to be denser but have worse power consumption characteristics against the TSMC process,meaning it might make more sense for their higher end chips to use TSMC.

Also,having their larger volume chips on GF will probably help with the WSA they have with GF too.

There is also one other consideration - Apple. Apple is buying up lots of capacity from TSMC for 16NM,which probably means both Nvidia and AMD will be struggling for capacity for a few months. GF/Samsung already have a working process and as TSMC gets more of the orders from Apple,it means AMD can use that capacity to get GPUs out quicker.

This is the thing - cards like the GTX750TI were boring for most of us. However,the GPU started the rot at AMD in the last 18 months with regards to marketshare,as it displaced AMD from quite a few laptops and OEM desktops,so I expect this 120MM2 part might be a pipecleaner like the HD4770 was,but also to try and get back marketshare in prebuilt PCs from Nvidia,especially if the competing lower end Nvidia GPUs are not being released for a while.

They would still have to ship the GPUs from NY to Malaysia or China even if GF chips were low end GDDR5(x). I don't see that being something that AMD would willingly ever do. High volume CPUs? Sure. Not GPUs.

No it isn't. You have no idea. No-one outside of the industry that's not had 'identical' (not really identical) test chips made on Samsung LPP and TSMC + knows for sure. But it's extremely unlikely however that what you state is true. The latter is a high power process, able to support high voltages. LPP is still a low power process, just supporting various features that LPE doesn't, and slightly higher voltages than its sister process. Almost by definition it will leak less and consume less power. You're thinking of iPhone chip tests done on LPE vs TSMC's bog standard one, which may well be better ... but there may be other factors at play there too. There's never been any suggestion of any GPUs being fabbed on either LPE or TSMC's basic process, so the iPhone tests are moot.

Density is slightly in Samsung / GF's favour, but I don't think it's really an issue that either side's likely to exploit. Yields and costs I'd expect to be massively in favour of Samsung.

Samsung have vastly more capacity at 14nm anyway ... I think they're soon to have 4 different fabs churning them out (AFAIK 3 are currently online). They had a head start on the new node, and have ramped really quickly.

I think you'll find that AMD launch pretty much everything mid '16 ... there won't be a big wait into later '16.

 

[GoogalyMoogaly]

Let's just hope that neither side have any issues with HBM2 stock. Lower stock levels could lead to increased prices or even just the obvious lack of stock of the GPUs. Both of which hurt us as consumers.
Like RAM and SSD prices, let's hope the market is flooded with HBM2 stock (and thus GPUs) so prices can be as low as possible.

 

[bru]

I can only assume that a lot of people didn't read the Anardtech article.

The card shown was a small part and most likely used GDDR5.

In any case, the GPU RTG showed off was a small GPU. And while Raja’s hand is hardly a scientifically accurate basis for size comparisons, if I had to guess I would wager it’s a bit smaller than RTG’s 28nm Cape Verde GPU or NVIDIA’s GK107 GPU, which is to say that it’s likely smaller than 120mm2. This is clearly meant to be RTG’s low-end GPU, and given the evolving state of FinFET yields, I wouldn’t be surprised if this was the very first GPU design they got back from Global Foundries as its size makes it comparable to current high-end FinFET-based SoCs. In that case, it could very well also be that it will be the first GPU we see in mid-2016, though that’s just supposition on my part.

First and foremost, Polaris will encompass both GDDR5 and High Bandwidth Memory (HBM) products. Where the line will be drawn has not been disclosed, but keeping in mind that HBM is still a newer technology, it’s reasonable to expect that we’ll only see HBM on higher-end parts. Meanwhile the rest of the Polaris lineup will continue to use GDDR5, something that is not surprising given the lesser bandwidth needs of lower-end parts and the greater cost sensitivity.

Meanwhile RTG has also disclosed that the first Polaris parts are GDDR5 based. Going hand-in-hand with what I mentioned earlier about RTG’s Polaris demonstration, it seems likely that this means we’ll see the lower-end Polaris parts first, with high-end parts to follow.

 

[tommybhoy]

I can't be the only one that gets fed up reading a group slating a technology or feature at one point and then 6 months later posting links and praising the other side for pushing the future tech when they do the same thing.

It's looking like only seeing what you want to, swings n roundabouts, remember 290X being banded every other thread about it being a throttling thirsty nightmare, TX came out and never heard it again-its a 2 way street and didn't hear much complaint from your end.

Each to there own I guess but considering the size that groups shrunk to, might get a better bite aiming at the other side.

 

[CAT-THE-FIFTH]

They would still have to ship the GPUs from NY to Malaysia or China even if GF chips were low end GDDR5(x). I don't see that being something that AMD would willingly ever do. High volume CPUs? Sure. Not GPUs.

But the chips are GDDR5 and the article said it was a GF/Samsung made chip.

You also need to consider part of the reason AMD as a whole was making losses was due to the WSA,and not meeting pre-allocated targets they made with GF,and hence they either needed to pay a penalty or over-produce CPUs which meant probably lower losses as inventory write downs,but they were losses anyway.

By shifting the console SOCs to GF,and by making large volume GPU parts at GF/Samsung it means they can fulfil the WSA much easier now.

The other aspect is that also with Apple pushing as much of their own chip production to TSMC as possible,Samsung is only being used to fulfil orders that TSMC cannot make. The latest iPhone chips are make at TSMC and Samsung and the iPad Pro chips at TSMC. Before,it was exclusively Samsung and as time progresses,Apple is trying to reduce its usage of Samsung as possible.

Remember,Samsung was producing functional 14NM chips in volume BEFORE TSMC managed to do so,so it is probably a more mature process at this time with regards to volume and has probably more free volume available too.

Plus,it would not surprise me if AMD gets some tax incentives in using a US based fab to make chips too.

This means AMD has a chance to use Samsung much earlier than TSMC,which gives it a leg up on bringing its products to market. Then add the fact the Samsung process is denser,and that Samsung is also an HBM2 producer too and is involved in a legal battle with Nvidia,it pretty much seems an advantageous time to use them.

It also gives AMD,more experience with using the process before they make Zen on it. AFAIK,they will be using the same,if not a derived process to make it.

 

[bru]

I very much doubt that the legal battle between Samsung and NVidia will have any bearing on Samsung supplying NVidia memory, just as the Samsung vs Apple didn't effect Samsung supplying Apple parts for their iphones.

 

[CAT-THE-FIFTH]

I very much doubt that the legal battle between Samsung and NVidia will have any bearing on Samsung supplying NVidia memory, just as the Samsung vs Apple didn't effect Samsung supplying Apple parts for their iphones.

The Apple contracts were worth billions to Samsung though and they could not realistically beat Apple in legal terms too,so did not even have any leverage. Plus notice how Apple has started reducing its reliance on Samsung?? The plant in NY was made mostly for supplying chips for Apple!! So in reality it did have an effect on Samsung,as they are moving more and more production and more of its parts manufacture to other companies.

However,in the case with Nvidia,they convincing won the first two rounds,so probably have degree of leverage already - I don't think HBM2 availability will be huge now for either Hynix or Samsung for a while since GDDR5 is being used in the volume Polaris part,but AMD is meant to be using it for some of their APUs too,which might potentially make them quite a big customer depending on how well Zen does,but then Nvidia does sell more GPUs than AMD.

They won't do the totally foolish thing and not supply Nvidia with HBM2 as they probably have already made the relevant contracts,but if there is not much floating around and AMD can take both Hynix and extra Samsung HBM2 allocations for a few months,I don't see why they won't help AMD a bit,in light of what has recently happened.

It not only means they do one at Hynix who are a competitor,but also it is still a sale for them,whether Nvidia or AMD buys it. What is Nvidia going to do since Samsung and Hynix are the two major RAM companies out there??

Plus if the rumours of Nvidia rejecting Samsung for chip production are true,wouldn't that make AMD a bigger customer for their tech currently?

They couldn't do that with Apple,since nobody else would be buying the custom parts they made for them. HBM2 is more a commodity part than a specialist part made for one company.

Plus it will be interesting to see what company gets the Polaris mobile dGPUs first.

I also believe Nvidia might get first dibs at 16NM a bit quicker than AMD will. IIRC,there was some info about them buying 16NM capacity before any mention of that regarding AMD,and I think they will be larger parts too,than what AMD is aiming for with the initial Polaris releases.

 

[bru]

The Apple contracts were worth billions to Samsung though

And you think supplying memory for a GPU series wouldn't be worth a lot as well?

 

[CAT-THE-FIFTH]

But the chips are GDDR5 and the article said it was a GF/Samsung made chip.

You also need to consider part of the reason AMD as a whole was making losses was due to the WSA,and not meeting pre-allocated targets they made with GF,and hence they either needed to pay a penalty or over-produce CPUs which meant probably lower losses as inventory write downs,but they were losses anyway.

By shifting the console SOCs to GF,and by making large volume GPU parts at GF/Samsung it means they can fulfil the WSA much easier now.

The other aspect is that also with Apple pushing as much of their own chip production to TSMC as possible,Samsung is only being used to fulfil orders that TSMC cannot make. The latest iPhone chips are make at TSMC and Samsung and the iPad Pro chips at TSMC. Before,it was exclusively Samsung and as time progresses,Apple is trying to reduce its usage of Samsung as possible.

Remember,Samsung was producing functional 14NM chips in volume BEFORE TSMC managed to do so,so it is probably a more mature process at this time with regards to volume and has probably more free volume available too.

Plus,it would not surprise me if AMD gets some tax incentives in using a US based fab to make chips too.

This means AMD has a chance to use Samsung much earlier than TSMC,which gives it a leg up on bringing its products to market. Then add the fact the Samsung process is denser,and that Samsung is also an HBM2 producer too and is involved in a legal battle with Nvidia,it pretty much seems an advantageous time to use them.

It also gives AMD,more experience with using the process before they make Zen on it. AFAIK,they will be using the same,if not a derived process to make it.

I very much doubt that the legal battle between Samsung and NVidia will have any bearing on Samsung supplying NVidia memory, just as the Samsung vs Apple didn't effect Samsung supplying Apple parts for their iphones.

The Apple contracts were worth billions to Samsung though and they could not realistically beat Apple in legal terms too,so did not even have any leverage. Plus notice how Apple has started reducing its reliance on Samsung?? The plant in NY was made mostly for supplying chips for Apple!! So in reality it did have an effect on Samsung,as they are moving more and more production and more of its parts manufacture to other companies.

However,in the case with Nvidia,they convincing won the first two rounds,so probably have degree of leverage already - I don't think HBM2 availability will be huge now for either Hynix or Samsung for a while since GDDR5 is being used in the volume Polaris part,but AMD is meant to be using it for some of their APUs too,which might potentially make them quite a big customer depending on how well Zen does,but then Nvidia does sell more GPUs than AMD.

They won't do the totally foolish thing and not supply Nvidia with HBM2 as they probably have already made the relevant contracts,but if there is not much floating around and AMD can take both Hynix and extra Samsung HBM2 allocations for a few months,I don't see why they won't help AMD a bit,in light of what has recently happened.

It not only means they do one at Hynix who are a competitor,but also it is still a sale for them,whether Nvidia or AMD buys it. What is Nvidia going to do since Samsung and Hynix are the two major RAM companies out there??

Plus if the rumours of Nvidia rejecting Samsung for chip production are true,wouldn't that make AMD a bigger customer for their tech currently?

They couldn't do that with Apple,since nobody else would be buying the custom parts they made for them. HBM2 is more a commodity part than a specialist part made for one company.

Plus it will be interesting to see what company gets the Polaris mobile dGPUs first.

I also believe Nvidia might get first dibs at 16NM a bit quicker than AMD will. IIRC,there was some info about them buying 16NM capacity before any mention of that regarding AMD,and I think they will be larger parts too,than what AMD is aiming for with the initial Polaris releases.

And you think supplying memory for a GPU series wouldn't be worth a lot as well?

Read my previous answer in full first,as I added more information,not the black and white answer you want - you are twisting it to make it like I said Samsung would supply no HBM2 to Nvidia which is NOT WHAT I SAID. You implied that. I implied that Samsung would probably BE MORE LIABLE to help them out with EXTRA(take note of that word) capacity FOR A FEW MONTHS(not a year or years just in case you thought it was that),considering recent events:
1.)Nvidia rejected them(supposedly) for making chips
2.)Nvidia took Samsung to court
3.)Samsung already has probably made contracts with Nvidia which they cannot break ,otherwise they will be sued
4.)Hynix is a largish competitor who is struggling to produce enough HBM it appears. So its a good chance to troll a competitor in that case.


Plus,no I don't think HBM2 contracts will be worth billions this year. GDDR5,DDR4 and GDDR4 will be worth much more. HBM2 is low volume and is probably only going to be for a few high end cards.

In terms of battle lines,I still expect that GDDR5 based cards will still generate most of the revenue for Nvidia this year and the same applies for AMD.

See any other uses for HBM/HBM2 this year yet...nope.

If the contracts were worth billions this year,then that would be the ENTIRE lines from top to bottom for Nvidia and AMD. Its apparent its not going to happen. If Hynix who are probably producing more HBM and HBM2 than Samsung will this year,can barely supply enough for AMD,then that basically tells you volume is not enough this year.

The thing is if Samsung has been peed off by Nvidia,then what happens to any extra HBM2 allocations outside what they signed for?? Why wouldn't they take extra allocations on,if AMD can support them??

Would Nvidia respond saying they would never use Samsung again?? I think not.

Neither AMD or Nvidia can really dictate to Hynix or Samsung,especially after Elpida went bankrupt.

AMD only got first dibs on Hynix allocations since it helped pay for development in the first place.

OFC,this is all assuming that Samsung can produce enough HBM2 to even fulfil whatever contracts Nvidia has signed with them!

Edit!!

I will keep it at that,otherwise we will go round in circles,which is what generally happens here.

 

[Scougar]

I can only assume that a lot of people didn't read the Anardtech article.

The card shown was a small part and most likely used GDDR5.

But what was THIS in the demo? Sure, the first parts might be HBM based, but what is the demo version. Wording is key.

What I'm trying to say is, the power efficiency might appear better than it really is IF HBM is used in the demo card. If it isn't, and GDDR5 is used, then awesome, it means that even more power can be saved

 

[CAT-THE-FIFTH]

But what was THIS in the demo? Sure, the first parts might be HBM based, but what is the demo version. Wording is key.

What I'm trying to say is, the power efficiency might appear better than it really is IF HBM is used in the demo card. If it isn't, and GDDR5 is used, then awesome, it means that even more power can be saved

True,but the higher end cards might have more non-gaming features like additional DP performance,which will affect performance/watt and HBM2 will enable reductions in power consumption to help negate that. Plus it also shrinks the memory controller in size,freeing up more space for other functionality.

So,unless AMD and Nvidia do produce more gaming orientated cards with HBM2,then core efficiency improvements will still be the more important factor.

With AMD focusing on efficiency with Polaris,and with Pascal improving compute performance(a partial return to hardware scheduling like with Fermi or GCN??),i do wonder if both companies will be more evenly matched this time in both compute and efficiency.

I also hope for the sake of AMD,they actually do a decent launch for once - every launch since the R9 290 series has been a mini-disaster IMHO.

 

[CAT-THE-FIFTH]

This is quite interesting and is referred to in both the Anandtech article and the one from Hardware.fr:

As a result a lot of the disclosed GCN 4 key features point to improving throughput of GCN. Improved shader efficiency is somewhat self-explanatory in that regard. At the same time RTG is also disclosing that we will be seeing some kind of hardware scheduler in GCN 4 along with instruction pre-fetch capabilities, which again should help them improve throughput in ways to be determined. Meanwhile in an improvement for the GPU front-end, GCN 4 will be adding a primitive discard accelerator – and again we don’t have any further details than this – but it should help the architecture clamp down on getting rid of unseen geometry. Finally, GCN 4 will also include a newer generation of RTG’s memory compression technology, this coming just one generation after it was most recently (and significantly) improved for the third generation GCN architecture (GCN 1.2).

Finally, AMD mentions a Primitive Discard Accelerator or an ejection system triangles masked the rendering pipeline. Remember, statistically, about half of the triangles of an object turning their backs to the camera and may be ejected from the rendering soon as this state is confirmed. Able to quickly make geometry can boost performance in real situations. Currently, the Radeon geometry engines are not able to perform this task faster than rendering a triangle, unlike GeForce who benefit to stand out in some scenes, especially when tessellation generates many hidden triangles. With Polaris, AMD should finally fill the gap, probably doubling the number of ejection motor engines primitive rasterization (Nvidia opted for a different approach by decentralizing some of the geometry processing but we do not expect that AMD follows this route).

So realworld tesellation performance should see a decent uplift too.

 

[D.P.]

Remember that AMD have exclusive priority access to HBM2, due to them helping develop the technology.

NVIDIA will only have sloppy seconds, hence may choose to adopt a more relaxed release schedule.

The most recent rumours had Nvidia with the priority access to HBM2. In bussiness money talks, whomever stumps up the cash wins

 

[YB2ndBEST]

Definitely what they're currently showing off is intended for mobile/ laptop use. Not the high end. This was using gddr5+

http://www.pcworld.com/article/3018...unces-new-4th-gen-polaris-graphics-cores.html

 

[Dave2150]

The most recent rumours had Nvidia with the priority access to HBM2. In bussiness money talks, whomever stumps up the cash wins

I heard a rumour that NVIDIA were going to swap from producing GPU's and start producing lunch boxes.

I can't provide a source, neither can you, such is life.