[Rumor - WCCFTech] AMD Arctic Islands 400 Series Set To Launch In Summer of 2016

[Techhog]

It depends by what you mean by replace the 390x and Fury. Performance wise, I'd hope so, but it terms of absolute placement I don't see anything like Fury coming out for awhile.

I really expect the two new GPUs to be replacements for Tahiti and Pitcairn. I'm not sure we'll see an analog to Cape Verde or if that ~100mm^2 mainstream card will just get pushed to a 14nm APU.
Something along the lines of a 10B transistor ~350mm^2 HBM2 big chip, and a 6-7B transistor 200-250mm^2 smaller chip. They'll outperform Hawaii and Fury by a decent margin owing to the newer architecture and greater number of transistors, but not by a huge amount. I would guess max 50%. I expect will have to wait a least a couple years before we see a 600mm^2 replacement for Fiji in their stack, hopefully less than the 3.5 years between Tahiti and Fiji.

Pure speculation on my part with no credible source though, so take it for what it's worth.

This is the most ridiculously pessimistic statement I've read unless I'm completely misreading it, especially since Tonga already exists.

 

[MrTeal]

This is the most ridiculously pessimistic statement I've read unless I'm completely misreading it, especially since Tonga already exists.

Well, Tonga really just is Tahiti with some extra transistors used to bring in the GCN1.2 feature set.

What I'm saying (to crib nvidia's nomenclature) is that we'll see AI104 (Tahiti size and launch prices) and AI106 (Pitcairn size and launch prices). The big dies (AI100) won't come for awhile, whether we get one in the 400 series with a Fury branded card, or they wait for the 500 series.

 

[dark zero]

Ok.... I'll give my estimates
- Mid AND High Tier: HBM1 and 2 comes at August 2016
- Low Tier: GDDR5 comes at November 2016 (Black Friday)
- God Tier: 2X HBM2 card (Dual GPU) at May 2017

 

[RussianSensation]

They will replace the R9 390/390X, and all Fiji cards.

For sub $300 in 2016 Tonga will be fine, R9 380X at $200-$220 and R9 380 at $150-$180.

Wow, a lot of you guys think they'll drop Hawaii and all Fiji chips entirely and just keep Tonga. Interesting.

BTW, the prices you listed for 380X and 380 are basically already there right now. Thus, it would make no sense for AMD's new 2016 line-up to include Tonga at $150-180 and $200-220. That's simply too slow level of performance for those prices.

Everything is going to need to shift 30-70%+ in performance over the next 2 years since this leap is akin to HD6000 -> HD7000 series. Performance increases and price/performance increases of that generation were huge over the course of 2012-2013. Once next gen games arrived, HD7000 pulled even farther away from HD6000 series.

As I mentioned earlier, the issue with Tonga is that the die size is barely smaller than Hawaii's but the performance is 50% worse. That means if AMD were to spend any $ at all shrinking Tonga or Hawaii, it makes sense to discard Tonga completely and use Hawaii. Better yet, throw both away and replace them with a brand new 16nm chip entirely. Shrink Fiji to 16nm to occupy the mid/mid high-end and bigger die AI with HBM2 for flagship $500+ cards.

I honestly think Tonga and Hawaii are going to be way too inefficient for next gen even if they are shrunk to 16nm. The perf/watt differences between these cards are the competition are too large to keep reusing them. A shrunken Nano/Fury X could be a power-house mid-range $349-399 card though.

This is the most ridiculously pessimistic statement I've read unless I'm completely misreading it, especially since Tonga already exists.

I am not entirely sure what he even meant because AMD already replaced Pitcairn with Tonga in the $140-230 price range and they replaced Tahiti 7950/7970 with $280-380 Hawaii 390/390X. Therefore, the replacement chips would have to be displacing Tonga and Hawaii not Tahiti and Pitcairn.

 

[MrTeal]

I am not entirely sure what he even meant because AMD already replaced Pitcairn with Tonga in the $140-230 price range and they replaced Tahiti 7950/7970 with $280-380 Hawaii 390/390X. Therefore, the replacement chips would have to be displacing Tonga and Hawaii not Tahiti and Pitcairn.

As I clarified above, I don't think we're going to get a big die chip out of AMD for awhile. The launch will be similar to 28nm, where we have a GCN 2.0 300-350nm chip similar to Tahiti, and a smaller one similar to Pitcairn. We won't see 16nm replacements for the Hawaii and Fiji GPUs for a couple years until they get more experience with the node and costs and yields improve. They'll obviously replace Fiji and Hawaii in the product stack and perform better, but they won't be the massive increases in performance we'd hope for, and (if I am correct on the sizes of the new dies) there definitely wouldn't be room for two new chips on top of a shrunk Fiji.

I'd expect AMD to return to the 400+mm^2 die space with a Hawaii replacement as part of the 500 series a year to 18 months after the 400 series cards launch.

 

[GodisanAtheist]

Regardless of what is released in terms of die sizes, we really need to get 390x performance in the $200 segment, Fury X performance in the ~$350 range and Fury X +30-50% at $500.

I think that's a realistic expectation from the first round of next gen cards. AMD can easily do this and more with two dies and a bevy of harvested skus.

 

[looncraz]

As I clarified above, I don't think we're going to get a big die chip out of AMD for awhile. The launch will be similar to 28nm, where we have a GCN 2.0 300-350nm chip similar to Tahiti, and a smaller one similar to Pitcairn. We won't see 16nm replacements for the Hawaii and Fiji GPUs for a couple years until they get more experience with the node and costs and yields improve. They'll obviously replace Fiji and Hawaii in the product stack and perform better, but they won't be the massive increases in performance we'd hope for, and (if I am correct on the sizes of the new dies) there definitely wouldn't be room for two new chips on top of a shrunk Fiji.

I'd expect AMD to return to the 400+mm^2 die space with a Hawaii replacement as part of the 500 series a year to 18 months after the 400 series cards launch.

I'd agree if AMD could actually afford to do so. If the new GCN revision is dramatically faster, then we might see it, but AMD can't afford to fall that far behind nVidia and may take their shot with big dies that they cut down even for the top-tier GPU (nVidia did this before, IIRC).

They are also generally limited by the interposer. Fiji is as big as it can possibly get without using multi-patterning on the interposer, which dramatically increases costs. AMD will absolutely stay below that, and so, too, may nVidia.

Have there been any rumors about GCN 2.0's design? We know Fiji effectively maxed out GCN 1.x's capabilities.

 

[Techhog]

Well, Tonga really just is Tahiti with some extra transistors used to bring in the GCN1.2 feature set.

What I'm saying (to crib nvidia's nomenclature) is that we'll see AI104 (Tahiti size and launch prices) and AI106 (Pitcairn size and launch prices). The big dies (AI100) won't come for awhile, whether we get one in the 400 series with a Fury branded card, or they wait for the 500 series.

That's a decent theory, but it kinda falls apart because Tahiti was originally intended as the true GCN 1.0 flagship and only became a midrange part because AMD made faster parts. Also, as RS said, Tonga is the the replacement for Pitcairn (and is more than just extra transistors stuffed into an existing die, but I digress) and Hawaii was the replacement for Tahiti, not a big die that was held back from the initial release.

 

[Techhog]

Regardless of what is released in terms of die sizes, we really need to get 390x performance in the $200 segment, Fury X performance in the ~$350 range and Fury X +30-50% at $500.

I think that's a realistic expectation from the first round of next gen cards. AMD can easily do this and more with two dies and a bevy of harvested skus.

The sad thing is that some aftermarket 980 Tis are already nearly 30% faster than the Fury X...

 

[GodisanAtheist]

The sad thing is that some aftermarket 980 Tis are already nearly 30% faster than the Fury X...

-Yeah, it looks like 20-25% faster with a CLC for ~$700 based on some of the recent reviews.

However with a node shrink, HBM memory, a much smaller form factor, energy and heat savings, a lower $ and the promise of future support 30%+ faster than a Fury X would be a fine place for a new mid-size high end GPU to land.

 

[MrTeal]

What do you think it would take to reach Fury X + 50% with GCN2.0, and would such a card retail for $500? When the 7970 came out, it was 30-40% faster than a 6970, but came in at 4.3M transistors on a 352mm^2 die vs 2.6M on a 389mm^2 die. Per transistor performance really didn't go up, in fact with launch titles it went down vs VLIW4.
Even GM104 Maxwell (in the 980) was only 60% faster than GK104 in the 770, despite almost 50% more transistors and a substantial clock speed advantage.

I'm hopeful that GCN2.0 will give a decent boost in per transistor performance, but that really hasn't been the trend. Now, if AMD does what nVidia does and cuts out a lot of compute to make a pure gaming card, there might be more there. For reference the 780Ti was ~35% faster than a 770 and had 50% more transistors. Let's say AMD does a great job, and get 20% better performance per transistor than Fiji. They'd still need 25% more transistors than Fiji to reach 50%. With 2x the density at 16nm, that would be about 372mm^2 for a die size.

That's a pretty big die on a new process, along with the costs associated with HBM2. I wouldn't be surprised is they got to 30-50% faster than Fiji, but I would be pretty shocked if they sold it for less than they're selling Fury X for now. This is the same company that with Tahiti made a card with a similar performance improvement but charged 57% more money for it.

 

[3DVagabond]

This is the same company that with Tahiti made a card with a similar performance improvement but charged 57% more money for it.

They started with bargain pricing on the 5870. It took them a couple of generations and additional sku's stacked on the top to recover from that pricing. With that said, it's obvious that AMD is increasing their ASP.

 

[AtenRa]

Well for 2016 im expecting,

a ~350mm2 die with HBM 2.0, + 30-40% faster than Fiji at $650-800 price range to replace Fury, Fury nano and Fury X.

And a ~250mm2 die with GDDR5, + 20-30% faster than Hawaii at $350-500 price range to replace R9 390/390X.

Perhaps we could see this smaller die (cut down CUs) to replace Tonga at the $150-200 range using only half the memory capacity and memory bandwidth from the SKUs above.

 

[beginner99]

Well for 2016 im expecting,

a ~350mm2 die with HBM 2.0, + 30-40% faster than Fiji at $650-800 price range to replace Fury, Fury nano and Fury X.

And a ~250mm2 die with GDDR5, + 20-30% faster than Hawaii at $350-500 price range to replace R9 390/390X.

Perhaps we could see this smaller die (cut down CUs) to replace Tonga at the $150-200 range using only half the memory capacity and memory bandwidth from the SKUs above.

Wouldn't a low end chip make more sense? Fiji can easily be moved down to middle be it current version or a die shrunk one. So 1 new chip with HBM2 and a low-end one with GDDR5. The low end-one would not be that slow but more importantly have newer features compared to current low-end (GCN 1.0) and very power efficient. For laptops.

 

[maddie]

What about the technical possibility of using the interposer to bind two GPU die?
This should be much better than Xfire and allow a top end solution with two smaller die.
If possible, this could allow a very full lineup with only the two new die promised by AMD.

 

[Head1985]

Current Fiji should be on 16nm about 300mm2

TSMC’s 16FF+ (FinFET Plus) technology can provide above 65 percent higher speed, around 2 times the density, or 70 percent less power than its 28HPM technology.

I am expecting from NV 80% more performance from GTX970/980 successors(its always same with new node) so about GTX1070 20-30% above TITANX and GTX1080 40%.That GPU should be about 300-320mm2

So if AMD want be copmpetetive they need update current FIJI with GCN2.If they wider front-end from 4pipelines to 8(and increase 64 rops to 128rops) they shold be around 320-330mm2 and that GPU should be 30-40% above TITANX.

Both cards will be around 400-550USD.

 

[Headfoot]

What about the technical possibility of using the interposer to bind two GPU die?
This should be much better than Xfire and allow a top end solution with two smaller die.
If possible, this could allow a very full lineup with only the two new die promised by AMD.

I definitely think this is a very interesting possibility, but I doubt it will come to fruition. Maybe if AMD had more "blue sky" R&D funds for that sort of thing (like the way Intel had Larabee cooking on a back burner eventually turning into Xeon Phi).

2 dies on an interposer would be essentially like a super MCM. If they made them out-of-the-box ready to link to each other like that, they could actually pursue the "many small die" strategy of yore, stacking 1 to 4 on an interposer. It's all probably a pipe dream though. A cool idea in concept

 

[AtenRa]

Wouldn't a low end chip make more sense? Fiji can easily be moved down to middle be it current version or a die shrunk one. So 1 new chip with HBM2 and a low-end one with GDDR5. The low end-one would not be that slow but more importantly have newer features compared to current low-end (GCN 1.0) and very power efficient. For laptops.

I dont have all the data i need as of now but i believe the $300 to $650 segment is the range that leaves the higher profits and have enough volume. That is what AMD and NVIDIA will aim for with the first 16nm GPUs.

If true, AMD and NVIDIA are going to bring new GPUs for that segment first. Fiji will be outdated for the next gen 16nm GPUs of $300-500 segment.

 

[maddie]

I definitely think this is a very interesting possibility, but I doubt it will come to fruition. Maybe if AMD had more "blue sky" R&D funds for that sort of thing (like the way Intel had Larabee cooking on a back burner eventually turning into Xeon Phi).

2 dies on an interposer would be essentially like a super MCM. If they made them out-of-the-box ready to link to each other like that, they could actually pursue the "many small die" strategy of yore, stacking 1 to 4 on an interposer. It's all probably a pipe dream though. A cool idea in concept

Going even further down the rabbit hole.


Die A optimized for gaming with low DP flops.
Die B optimized for HPC with full DP flops.

Use multiple die to achieve desired performance target.
Advantages:
1) For a given total GPU area, multiple small die will be cheaper than a monolithic one.
2) The possibility to go past the normal 600mm^2 limit for a big die GPU.
3) Only 2 designs needed and each optimised for the market segment targeted.
4) Extremely high shader count running at lower speeds to reduce power
consumption [using the same design philosophy as HBM vs DDR5] + [dependent on point 2]
5) Lower inventory and production costs [no mask changes]

Disadvantages:
6) Power probably rises for a given level of performance. [ might be negated by point 4]
7) Introduces latency [GPU designers claim that higher latency can be accommodated]

 

[jpiniero]

I dont have all the data i need as of now but i believe the $300 to $650 segment is the range that leaves the higher profits and have enough volume. That is what AMD and NVIDIA will aim for with the first 16nm GPUs.

At least for nVidia, it's going to be Titan/HPC first. Given the performance increase at a given price point is not going to be that much, it probably makes sense to go top to bottom.

Edit: And 28nm Fiji would definitely be competitive from a performance perspective in the $300-$500 range. Margins would suck although it'd be cheaper than most alternatives.

 

[3DVagabond]

At least for nVidia, it's going to be Titan/HPC first. Given the performance increase at a given price point is not going to be that much, it probably makes sense to go top to bottom.

Edit: And 28nm Fiji would definitely be competitive from a performance perspective in the $300-$500 range. Margins would suck although it'd be cheaper than most alternatives.

We don't really know how much it costs to produce Fiji. I would imagine it's less today than when first released. Especially considering the TSV was just going into volume production at the same time. Add to that reduced price for HBM1, better yields for the actual GPU, etc...

As far as shrinking it, which I thought might be a feasibility. I've since read that since it's a planer transistor design and the new modes are finFET it's not really possible without a major redesign anyway. So, I don't see them keeping such a big chip in production when it's obvious a design on the new process would be better.

 

[RussianSensation]

So if AMD want be copmpetetive they need update current FIJI with GCN2.If they wider front-end from 4pipelines to 8(and increase 64 rops to 128rops) they shold be around 320-330mm2 and that GPU should be 30-40% above TITANX.

Both cards will be around 400-550USD.

I expect with next gen for both firms to start pricing faster cards at $650. There is no point not to. The market has spoken and paid $550 for 7970 then $550 for 290X and now $650 for Fury X. At this point the general market isn't looking at price/performance or die sizes or where the next gen chip falls in that entire generation. If something is 20-30% faster than the last gen's flagship, they can price it at $650+. 680/980 are as much proof as they need that this works. Since AMD is no longer interested in price/performance, why would they price a card 30-40% faster than the 980Ti/Fury X for $400-550? The only way for that to happen is if the competitor outperforms them so dramatically that they are forced to lower prices.

I am going to be pessimistic this time after all the Fury X hype. I am just going to assume that AMD's 2016 flagship is 30-40% faster than Fury X and then in 2017 they'll release another card 30-40% faster and both of these will cost $650 to maximize profits. The old days of getting a next gen $650 card that's 70-80% faster than the last gen's flagship are pretty much history. From now on, it's probably more likely to expect more frequent releases with 30-40% faster performance every 12-15 months. Over the entire 2-2.5 years of one generation, they'll net the same 80-100% increases in performance, but now both firms will sell you 2 or even 3 flagships in the same generation by splitting the gen up into parts.

 

[3DVagabond]

I expect with next gen for both firms to start pricing faster cards at $650. There is no point not to. The market has spoken and paid $550 for 7970 then $550 for 290X and now $650 for Fury X. At this point the general market isn't looking at price/performance or die sizes or where the next gen chip falls in that entire generation. If something is 20-30% faster than the last gen's flagship, they can price it at $650+. 680/980 are as much proof as they need that this works. Since AMD is no longer interested in price/performance, why would they price a card 30-40% faster than the 980Ti/Fury X for $400-550? The only way for that to happen is if the competitor outperforms them so dramatically that they are forced to lower prices.

I am going to be pessimistic this time after all the Fury X hype. I am just going to assume that AMD's 2016 flagship is 30-40% faster than Fury X and then in 2017 they'll release another card 30-40% faster and both of these will cost $650 to maximize profits. The old days of getting a next gen $650 card that's 70-80% faster than the last gen's flagship are pretty much history. From now on, it's probably more likely to expect more frequent releases with 30-40% faster performance every 12-15 months. Over the entire 2-2.5 years of one generation, they'll net the same 80-100% increases in performance, but now both firms will sell you 2 or even 3 flagships in the same generation by splitting the gen up into parts.

I'll do you one better and if they have the hardware they will release The $650 card you predict and then 3 months later a $900 flagship. Since it's not $1000 the AMD faithful will be able to defend them as not price gouging like Titan.

 

[RussianSensation]

I'll do you one better and if they have the hardware they will release The $650 card you predict and then 3 months later a $900 flagship. Since it's not $1000 the AMD faithful will be able to defend them as not price gouging like Titan.

Let's use Tonga as an example. AMD said GCN2.0 on 16nm is aiming to deliver 2X the perf/watt of old GCN parts.

That means 380/380X's successors would be aiming for 174% / 192% on this chart:

There you go, it's theoretically enough for them to beat Fury/Fury X with a Tonga successor. Tonga is more or less this gen's low-end chip. (Tonga = low end, Hawaii = mid-range, Fiji = high-end this gen). That means in the theoretical Arctic Islands 16nm generation, Hawaii successor's performance also goes up 2X, Fury X's also 2X.

It stands to reason that a true mid-range next gen 16nm AI card should smoke Fury/Fury X based on AMD's own 2X perf/watt claims. Now, are they going to price those cards at old historical mid-range HD5850/5870/7850/7870 prices? Not a chance.

 

[3DVagabond]

Let's use Tonga as an example. AMD said GCN2.0 on 16nm is aiming to deliver 2X the perf/watt of old GCN parts.

That means 380/380X's successors would be aiming for 174% / 192% on this chart:

There you go, it's theoretically enough for them to beat Fury/Fury X with a Tonga successor. Tonga is more or less this gen's low-end chip. (Tonga = low end, Hawaii = mid-range, Fiji = high-end this gen). That means in the theoretical Arctic Islands 16nm generation, Hawaii successor's performance also goes up 2X, Fury X's also 2X.

It stands to reason that a true mid-range next gen 16nm AI card should smoke Fury/Fury X based on AMD's own 2X perf/watt claims. Now, are they going to price those cards at old historical mid-range HD5850/5870/7850/7870 prices? Not a chance.

If you assume by old they mean the latest Tonga chips. I'll bet they are comparing to Tahiti with 1/3 DP. Not a compute crippled chip.