AMD’s Revolutionary 14nm GPU Architecture

It seems like AMD will finally release an actual mobile videocard this year, I am really hyped for this.
Maybe this is finally the replacement for my 7970m crossfire. MXM users rejoice!


More info(just click on the pictures): http://www.computerbase.de/2016-01/amd-polaris-next-gen-radeon-14-nm-finfet/#bilder

 

LOL AMD. Make a decent CPU already. You use your rival's CPU on your bench system.

 

Intel too strong ma friend xD

 

I'll believe it when I see it. AMD Fury with HBM2 very disappointing. I have little faith in AMD.

 

Wasnt fury with HBM1 ?

Ye of little faith perhaps it is time we consumers get a little competitive products

 

Fury X was HBM Gen 1 on a 3 year old architecture. Lets see what they can do with a 14nm process and HBM2. I have faith.

 

GCN is just being revised again, you know. Sure it's a die shrink, but the basic process is there. Hence why I'm on a wait-and-see deal. That being said, with what I've heard of Pascal, and what I inferred from Maxwell compared to other architectures, AMD might be making it power-efficient and cool, while nVidia might be returning to a hotter architecture (regardless of its power efficiency).

We'll see what HBM2 brings to the table. Currently, HBM1 is nothing worth writing home about over GM200's default effective memory bandwidth, and is even less impressive near its maximum bandwidth (8000*384*1.15/8 = ~441GB/s effective bandwidth on 980Ti/Titan X cards at 8GHz memory clock; rather close to the already-near-limits HBM1 bandwidth of 512GB/s), and then games barely use memory bandwidth a lot anyway.

 

Fury was on HBM1, and in terms of power efficiency, Fiji architecture (the entire Fury line) was only 10W less efficient than Maxwell on the desktop - although, when you factor in that AMD excels in professional software, then it starts making sense where that TDP and SP's are actually going (and it does not appear that gaming usually benefits that much from increase in SP's... but pro software is another matter).

All in all, Polaris was slated to have about 2 to 2.5 times more performance per watt compared to Fiji (Fury line) architecture (without HBM being mentioned contributing to this efficiency... though if I'm not mistaken, the links also suggest that HBM 2 will be used for the entire Polaris line - whether this extends to laptops, I don't know... I so hope it does though... AMD brought HBM first to the table, so maybe they could bring it first to laptops).

As such, the links about Polaris indicate that this is a replacement for GCN, and an entirely new architecture, not a tweak like Fury was - though, I guess that you could say any new architecture is a tweak or a revised version of previous ones, as its more or less better.

As for memory bandwidth... it becomes important with very high resolutions such as 4K, and I would imagine Virtual Reality.

 

If you count core-for-core... maybe. I mean people overclocked 980Ti cards to 1450MHz on the core barely passing 111% of 250W TDP without custom vBIOSes (and being incapable of passing... is it 300W? In total? I can't remember how to add the power pins to check delivered wattage). Now Fiji had many more cores than GM200; over 4000 vs over 3000. But Fiji couldn't even handle GM200 properly (with the drivers given, anyway; I can only go by what's apparent now) and that means each Fiji core is less powerful than a maxwell core. AND they were at their limits of overclocking as well, and often drew near their max TDP. The Nano is a different story; being binned for efficiency, but it opted for a 175W hard limit, and the card often throttled down in certain demanding titles as a result. I won't count the Nano because that isn't "average" architecture performance. From what I see, AMD was nowhere near being nearly as efficient as Maxwell in terms of performance per watt, considering the cards' limits in TDP when overclocking was performed. I'd say it was a big improvement over Hawaii, for certain. Hawaii and Fiji have the same TDP on average, but Hawaii is only about 60% as strong as Fiji at both their limits... again with the same TDP.

Now this here is not exactly a good point. What you're trying to say is, the cards themselves are capable of varying types of calculations not seen on Team Green since the hot and power hungry Fermi architecture. And this is true, however it shouldn't be using those engines when the card is just gaming. If it is running those things while the card is gaming, and it's not contributing to the gaming that's going on, that's dumb, and they need to fix that. But as far as professional programs go, most of them much quicker jump at CUDA (to my dismay).

The architecture itself states it's GCN gen 4. Or more correctly, "4th generation Graphics Core Next CUs". GCN = Graphics Core Next. Hence, a revision. If they want to say it's GCN and it's done basically from scratch, more power to AMD. The results are what matters.

 

GCN and Maxwell shader counts can't be directly compared. What matters is the end result.

As for it being a GCN revision, that's just how AMD choose to name each iteration. Kepler was VERY similar to Fermi in most ways and could easily have been called Fermi 1.5. Don't get hung up on names, Nvidia and AMD do it very differently. AMD should probably just move away from the GCN name because people seem to be thinking they aren't doing anything but adding shaders and memory bandwidth which obviously is not the case.

Anyway this is good news. Nvidia's arrogance is at record highs which is bad for everyone.

 

I'm not thinking they are just adding shaders etc. I'm thinking they are doing NOTHING but developing more junk marketing media.

 

It does wonders for Apple, maybe AMD will also succeed!

 

Your thinking however doesn't seem supported by the evidence which demonstrates AMD Polaris GPU consuming almost 50% less power than desktop Maxwell GTX 950 and producing same frame rates (granted, these are presentational scores, so we cannot discount AMD tweaking the graphic settings a bit in terms of details - still, for conclusive analysis we will need to wait, and right not, preliminary findings are encouraging).

 

ROFL. They said GCN V2 would use less power and yet, it's still has rep for being power hungry and hot. AMD says a lot of BS.

 

nGreedia doesn't? Right. It's called marketing, it's part of the game, like it or not.

 

Except it does use less power. Compare Tahiti to Fiji, which has twice the amount of everything and uses either less power (Nano) or slightly more for a big performance gain. The perf/watt improved dramatically. The same is true for Hawaii.

But don't let these facts get in the way of your absurd rambling.

 

Everyone, stop fighting. Both AMD and nVidia are preparing big things for their new GPUs after a rather stagnant period of performance increase.

It's not even funny seeing these kinds of useless posts. We know you don't fancy AMD. Drop it if you don't have anything to add.

 

And yet according to WCCFTECH, the Polaris architecture will use exclusively 14nm FinFet:

[UPDATE 3 : 2016 January 8 12:23 ET] We’ve reached out to AMD and they have confirmed that Polaris GPUs are exclusively – Globalfoundries/Samsung – 14nm FinFET based, rather than a mix of both TSMC 16nm and Globalfoundries 14nm chips like some have reported.

 

if they can do an MXM variant and it works with Alienware laptops with MXM cards then i think we may have a serious contender to make me go to the red team!

 

Since when does Zymphad's opinion matter anyways, all he does is make BS claims and thinks nVidia is "be all, end all".
He has a nVidia logo for his avatar, that about tells you what you should expect from him.

 

If you guys start fighting again, I'm issuing infractions to everyone. There is already one warning in the thread from ryzeki.

 

Apparently AMD showed of Polaris at CES with a 950 equivalent drawing only ~38 W during a session of battlefront at high settings:

https://translate.google.se/transla...ikarkitekturen-polaris-pa-ces-2016&edit-text=

 

If AMD can get this in MXM, bye bye NVIDIA...

 

If that were the case, AMD would have played fair and pit Polaris vs a laptop chip since that's what Polaris is designed for. The fact they chose to use a 90W 950 desktop card which is slightly faster (5-15% depending) than the 50W 965M speaks volumes... And keep in mind, nVidia got the energy consumption down on Maxwell while on the same 28nm node. The gains from Pascal should be much more impressive.

I'll reserve judgment for final parts but I'm not really all that impressed with AMD basically pitting a chip they're focusing on laptops against a wasteful desktop card.

 

Who said that it's indeed a designed for laptop chip? As far as I understood, they say that it would be perfect for laptop use. I also see "but also power efficient and budget-oriented desktop systems". And Maxwell's efficiency is (mostly) gimmick - cut-down FP64; cut-down color channels (take a look what happens to anyone trying to upgrade to 9x0m with more than 6bit panel, i.e. M17x-R2 (8bit), 8740w DreamColor (10bit)); cut-down or more likely on-demand ultra fast power switching, which is nothing once you overclock and also induces instabilities in SLi machines. That's A LOT of cut-downs to get that "efficiency". So yeah, tell me more about that "glorious" thing called Maxwell.

 

I didn't say Maxwell was glorious, I said that nVidia managed to pull a rabbit out of its hat on the same node. I wouldn't underestimate Pascal, nVidia has had a lot of time to tweak the design. nVidia didn't really need to do much for Maxwell, it was just meant as a stopgap solution.

As for Polaris, its obvious that AMD is demoing a chip that is going to find its way into a laptop. The fact they didn't show the card but the article seems to insinuate that its a nano card without HBM tells me its just a PCIe MXM board. Either way, no sense in arguing over it until more details come out but I stand by my statement that if AMD is intending to take this chip back into the laptop sector, they should put it up against a comparable laptop chip (which it still beats in efficiency anyway, just not as drastically).

 

I won't underestimate Pascal (for that matter, I won't/wouldn't underestimate Polaris either and would prefer to wait for a lot more details - I also won't expect any 'miracles' either)... but we have had previous indications where power efficiency on Nvidia's side wasn't all that was cracked up to be.
And as 'triturbo' mentioned, Nvidia introduced 'cuts' in one way or another to get relative power efficiency as shown in Maxwell which quickly evaporates once you start with your own modifications.

If I am not mistaken (and do correct me if I am), AMD doesn't exactly gimp its gpu's in the same way Nvidia does, which is why they usually manage to heavily overtake Nvidia in professional software and why their power draw is higher (plus it would explain where the higher SP numbers are going).

Oddly enough, very few people seem to be taking pro software performance on AMD gpu's when comparing them to Nvidia... they usually focus on gaming, and that in itself is not a good comparison as it seems incomplete and biased.

Apart from that, this Polaris part that was demo-ed appears to be entry level, plus it was tested in a mini ITX solution - which might indicate a desktop grade GPU (but if it was anything like R9 Nano, I would imagine it would be a lot more powerful).
I think a Nano style gpu might be in the works... but it might come at a later time (with release of higher end GPU's - as I think the whole point of it was to provide Fiji level of performance [with relatively small loss], in a much smaller power envelope).

At this point, we do know that Polaris was mentioned as suitable for use in laptops, so the actual mobile versions will likely be even more power efficient and possibly faster.

It would be interesting to see a Nano style Polaris mobile gpu with HBM2.

 

I'm keeping my eye on AMD as well, especially for the mid range segment. If they can compete with Nvidia there, it may be well worth investing in their tech instead of Nvidia. Would be good to see AMD out there again, and maybe kick Nvidia's butt a bit.

 

That's what I'm hoping to see, but I'm holding out until the high-end mobile gpu's from AMD come (and I'll probably upgrade my ageing Acer 5930G then).

It would be nice if HBM 2 was part of AMD's high-end line-up.

 

AMD just dropped the price on the Fury Nano to 499. They've got something up their sleeves.

 

Trying to get back some of the market cap. NVIDIA's been eating it up.

 

So why only drop the price on the Nano? I wonder if they're going to bring a Polaris Nano card out... I'd be lying if I said I wouldn't be tempted to replace my 780 Ti with one.

 

A Polaris Nano gpu with HBM2 would likely be really good... especially if made in mobile form.

 

Good good news! Seems like AMD is going to release laptop versions of the next gen gpus first and then desktop versions to follow! All these in 6-7 months timeline. I'm dying to change my 7970M, so hopefully we'll have smth at mobile enthusiast level, rather than entry-middle class!
http://wccftech.com/amd-confirms-polaris-gpus-released-school-season-desktops-laptops/

 

If their naming scheme is like before, Polaris is a rather low-end model only competing with a 950. We'll wait and see what the higher end models do. Let's wait and hope.

 

I agree. It would have been much more interesting if they had released more info specific to Greenland.

 

They also mention an enthusiast part to replace fury. Who knows they might release a slightly chopped down version of this and stick it on mxm.

Sent from my SM-A500FU using Tapatalk