AMD Polaris GPU: detailed slides leaked

[kalniel]

But the problem is Nvidia has ripped a lot out to get there.

Good point and I think that's key - the 'best' cards were the ones that were very very dedicated to their cause, harming their longevity in the process most likely. Personally I've always preferred better all-round ability and hence AMD, but it will be interesting to see if they feel the need to specialise again at some point. I wouldn't put it past them to have a gen 2 product go down that route.

I do really like it when AMD back off a little bit and hit the sweet spot of speed vs efficiency as seen in Pitcairn and Nano. Hopefully Polaris will give them confidence to do more of the same.

[watercooled]

As a casual gamer I don't really get the top end cards but I'm really excited by this and whatever nvidia follow up with. Low power, smaller cards with the meat to play the top titles whenever you need it

Or, of course, we could see certain game developers exploit the extra performance as an excuse to have their games run like crap. Good old Wirth's law...

[scaryjim]

... I've always preferred better all-round ability and hence AMD, but it will be interesting to see if they feel the need to specialise again at some point. ...

If you look at compute performance between Hawaii and Fiji they're very similar despite Fiji having significantly more raw power - so AMD have *already* started trading compute for power efficiency. It makes me wonder if they're going to look at doing two lines of GPUs and have dedicated silicon for the professional and HPC cards, or if they're going to rely on efficiency gains for HSA to encourage people to go all AMD for workstations & HPC in the future.

Nvidia have already had a generation with a huge improvement in energy efficiency so I doubt they've got anywhere near as much to gain from going 14nnm finfet as AMD do. Nano's perf/watt is better than the GTX 980 already, and the figures I've run suggest that AMD are talking about 2.5x better perf/watt compared to Fiji, which means NVidia are going to have to pull the same efficiency improvements to keep up. I think AMD still having most of their cards at least a generation behind is somewhat masking where they're currently at in terms of perf/watt....

[CAT-THE-FIFTH]

Polaris 10 and Polaris 11 are the name of the two new GPUs:

http://wccftech.com/amd-unveils-pola...#ixzz3xQyWdqsX

If you look at compute performance between Hawaii and Fiji they're very similar despite Fiji having significantly more raw power - so AMD have *already* started trading compute for power efficiency. It makes me wonder if they're going to look at doing two lines of GPUs and have dedicated silicon for the professional and HPC cards, or if they're going to rely on efficiency gains for HSA to encourage people to go all AMD for workstations & HPC in the future.

Nvidia have already had a generation with a huge improvement in energy efficiency so I doubt they've got anywhere near as much to gain from going 14nnm finfet as AMD do. Nano's perf/watt is better than the GTX 980 already, and the figures I've run suggest that AMD are talking about 2.5x better perf/watt compared to Fiji, which means NVidia are going to have to pull the same efficiency improvements to keep up. I think AMD still having most of their cards at least a generation behind is somewhat masking where they're currently at in terms of perf/watt....

Nvidia is using TSMC 16NM and AMD GF/Samsung 14NM LPP. Both use Finfets and the 14NM LPE process is meant to be around 10% denser than the TSMC 16NM process at the cost of higher power consumption. However,14NM LPP is meant to improve power consumption so it could be that AMD not only has a time to market advantage but a process node density advantage this generation over Nvidia!

[kalniel]

If that's right they're going for one very big and one very small to start with.. not my preferred middle boo.

[CAT-THE-FIFTH]

If that's right they're going for one very big and one very small to start with.. not my preferred middle boo.

I was talking about it on another forum - Nvidia only have two GPUs,ie,the GM206 and GM204 serving the £100 to £450 range. So I suspect larger Polaris will fit the £250 to £400 range and smaller Polaris under £200.

I expect larger Polaris to be 300MM2 to 400MM2 which is around the same AMD aims for when using a node first and the smaller one is supposed to be around 120MM2,but the latter is a rough figure AFAIK.

So you might find the following structure:
1.)R9 480X ~ around £350 to £450
2.)R9 480 ~ £250 to £350
3.)R9 380 ~ £150 to £200
4.)R9 370 ~ £100 to £150

That's roughly where you are seeing the GTX950,GTX960,GTX970 and GTX980 pricing lying at and its done pretty well for Nvidia.

[kalniel]

You don't think they'll put larger polaris where Fiji is currently (ie a whole range above £400)?

[CAT-THE-FIFTH]

You don't think they'll put larger polaris where Fiji is currently (ie a whole range above £400)?

I doubt it as they said they had two products being released this year!

OFC,we are only basing this on the fact we saw AMD demonstrate a 120MM2 Polaris chip. OTH,it could only be a test chip and we might have a large chip,ie,200MM2 to 250MM2,as the lower end one,which might mean the one above it might be priced higher and closer to £500.

But if the lower chip is only 120MM2,I can't see how it would power a £200+ graphics card.

Also,I assuming they don't re-use their current products. If they were to re-use Fiji in the £250 to £400 range,it might be mean larger Polaris is above £400,but Fiji is a massive 600MM2 28NM die pushing the limits of what can be achieved with current interposer technology. So it might be the case it has no cost advantage over a 300MM2 to 400MM2 14NM LPP chip using an interposer,plus the fact the latter will probably require less power and less cooling meaning lower cost PCBs and coolers too.

The think is we have seen the last time Nvidia started hitting hard,ie, G80 series against HD2000 series,AMD prioritised time to market and price over absolute performance.

So,it wouldn't surprise me,to claw back sales,they might be doing the same.

Personally,I think Nvidia will have a higher performing part out before AMD this year,but most likely being a Titan part of some sort,but TBH that is what not AMD really needs IMHO.

[watercooled]

You don't think they'll put larger polaris where Fiji is currently (ie a whole range above £400)?

Personally I don't think that's the most likely scenario, taking into account yields for a huge die on a new node and the relative market share. The mid-range market ships massively more than the flagship range, and it can likely find a wider addressable market because of its better suitability to laptop use, so it would make complete sense to target that before worrying about the costly huge dies, but then again the flagships carry the halo product marketing status so they're not exactly unimportant.

It's also unusual for huge dies to be the first released on a new node in general, for any of AMD, Nvidia or Intel. Going on previous generations and the little we know about 14nm, my estimate of die size would be similar to CAT's but IMO closer to 300 than 400.

[kalniel]

I'm a bit worried they'd go

>£400: Large polaris
>£300: Rebranded Fiji
>£150: Rebranded tonga/Hawaii
<£150: Small polaris

[watercooled]

Fiji is still fairly new and especially under the Nano bin it's really quite efficient, so I don't think it's out of the question it would continue to sell, but I'm not sure where it would fit in. The huge die size likely means cost is high even for 28nm so I can't see it selling too cheaply, but a much smaller ~300mm2 14nm GPU might end up completely outperforming it to the point of making it obsolete.

However, unless I've missed something, we really don't know anything about 'big Polaris' aside from its unofficially-confirmed existence and we don't know if it's HBM or GDDR.

I would have thought addressing the mid-high end and mobile markets is hugely important, something they might struggle to do effectively with one very small and one very large 14nm GPU, unless they have a cost-efficient way of addressing the mid-range market with the larger die. Also, if they 'only' manage to exceed Fiji's performance at lower power consumption/cost, it would arguably have less marketing value than a big performance jump, so who knows...

Edit: @CAT - what's the source for the small GPU being ~120mm? Because I agree that having the smaller chip larger than that would be another potential way of addressing the mid-range market, depending on yields.

[HalloweenJack]

a rebranded Fiji would work especially if they slide it down the scale where 2/4gb cards are the usual , and use HBM1 would mean a claim of HBM from medium to high end segment.

[CAT-THE-FIFTH]

The size is an estimate from Anandtech:

http://www.anandtech.com/show/9886/a...u-architecture

Regarding HBM2:

http://wccftech.com/amd-confirms-pol...bm2-and-gddr5/

[CAT-THE-FIFTH]

It looks like Polaris is targetting price points under $350:

http://wccftech.com/amd-polaris-arch...-minimum-spec/

[kalniel]

It looks like Polaris is targetting price points under $350:

http://wccftech.com/amd-polaris-arch...-minimum-spec/

Much lower in fact, given the point about $350 being too expensive. But I wonder if this is a mid-range chip with some specific hardware for VR so they can claim it's the minimum you need for VR without being an otherwise very powerful card?

[CAT-THE-FIFTH]

Much lower in fact, given the point about $350 being too expensive. But I wonder if this is a mid-range chip with some specific hardware for VR so they can claim it's the minimum you need for VR without being an otherwise very powerful card?

The R9 380 supports the same level of VR features as a Fury X,so I do wonder if it means R9 290/GTX970 level performance under £200?? That will be quite a decent jump over a R9 380 or a GTX960 IMHO!