Polaris 10 near 980Ti/Fury X performance for $300 and 175W TDP.
The wonders of FinFETs
http://www.tweaktown.com/news/51814/amd-polaris-10-gpu-performs-close-to-nvidia-geforce-gtx-980-ti/index.html
http://www.game-debate.com/news/?news=20006
Something something on suicide watch, let's all hope GP204/GTX1080 is near that performance as well, we can expect prices to go way down.
Unfortunately no info on diesize, but if Polaris 11 is around 120mm^2, Polaris 10 can't be be larger than 280mm^2 or AMD risks selling expensive dies for cheap.
>>54234216
So their next tech is only capable of competing with the current flagship? To be competitive ATI has to price it aggressively? Stop the presses!
>>54234296
This isn't their flagship, and Nvidia doesn't have a GPU to replace their current flagship until Q1, just like AMD.
Stop being retarded.
>not released
Thank fucking God AMD is focusing on small dies again like they did before.
Let Nvidia keep their 10% faster, 35% bigger and 40% power hungrier GPUs, it'll satisfy autists like Asian Ricer !!x2XTTmvcXn0
>>54234304
Polaris 10 is for high end desktops. If they are touting price/performance instead of simply greater performance there seems to be some big bolder words not getting printed trying to shout from between the lines which we printed.
>>54234260
One rumor said 232mm^2
>>54234362
According to AMD's own words, Polaris 10 is for 'mainstream' computers and high-end mobile(which would point really nicely at 980Ti performance) which if you take what we have today into account, be 950/370 performance.
And that obviously makes no sense, so either AMD is raising the bar on the 'mainstream' or we're being lied to.
Also, it's a relatively small die, it can't be highend unless their architecture is extremely efficient.
>>54234376
I've seen numbers on GP104 being around 320mm2, but not on Polaris 10, where did you get this?
>>54234216
perfect time for me to upgrade my GTX 670.
I'm glad I waited so long. 100$ cheaper than the 670 on launch too.
Hopefully AMD didn't neglect overclocking this time.
>>54234446
http://wccftech.com/amd-polaris-gpu-die-size-232-mm-2/
[spoiler]0.65 seconds in google[/spoiler]
>>54234465
Extremely impressive, 100mm^2 less than a GP104 but around the same performance.
AMD has got a winner here, lets hope their marketing doesn't fuck up.
[spoiler]It will[/spoiler]
>>54234463
They certainly neglect their linux drivers the fucking cunts
>>54234501
That's got nothing to do with overclocking, but I'll bite.
Polaris already has AMDGPU support starting from 4.16, what more do you need?
>>54234518
>4.16,
4.6*
>>54234410
>According to AMD's own words
Got a link to these words? Or are you considering Fury X2 to be the new flagships for high end destkop use rather than for VR?
>>54234463
That assumes AMD neglected overclocking instead of had to push their tech to the walls leaving little overclock headspace so that their cards could compete with the 9xx series.
I always believed in amd
>>54234552
>>54234518
>Not having a update since 2015
Yeah neglect fuck it
>>54234574
http://phoronix.com/scan.php?page=news_item&px=AMDGPU-Xorg-1.1-Driver
Mate, this was 19 days ago.
Will DX12 fix the SLI/Crossfire stuttering and compatibility issues?
Will dual Polaris 10 beat out Vega for price/performance?
>>54234570
So AMD is giving up on having a flagship competing with Nvidia in 2016 and betting on (and spreading out the technology) Vega is 2017?
>>54234623
No, Polaris 10 is going to compete with high-end Nvidia in 2016, Vega will compete with Pascal in 2017, do not let the "mainstream" disambiguation confuse you
>>54234623
Nvidia doesn't have a flagship in 2016 either, there's no way either can muster up the yields to produce 500mm2+ GPUs in high volume until later.
What I'm guessing based on diesize and leaks so far.
GP104/1080 will be around 5-10% faster than a 980Ti at around 170W average
Polaris10/490X will be neck to neck with 980Ti
at around 150W average.
I personally think AMD has the advantage here, their GPU is much smaller.
But these are top bins, the 490/1070 versions will be much more interesting since they'll be only 10% slower but much cheaper.
>>54234641
>Polaris 10 is going to compete with high-end Nvidia in 2016
Polaris is going to compete with high-end Nvidia from 2015 not pascal. If they are mainstream cards they are not going to compete with the high-end Nvidia in 2016 which will be based on Pascal rather than Maxwell.
>>54234666
Current iterations from Pascal are around highend 2015 Nvidia performance, but at a much lower wattage.
Same goes for AMD, neither company expects large performance gains until 2017 until GP100 and Vega 10 release.
>>54234493
>implying that'll happen
Nvidia have the process advantage (TSMC's 16nm is superior) and the money advantage. No way AMD beats them that badly.
>>54234342
Is this the return of Evergreen and RV700 but without Nvidia making jet engines?
GOD BLESS AMD
>>54234692
>(TSMC's 16nm is superior)
Hahaha.
Now you'll post me a anandtech paragraph about a 7 month old LPE process running on a 2W SoC as evidence, right?
>>54234692
Both die sizes are already confirmed with a 10% error.
I don't know what you're up to, but GloFo is proven to be denser.
>>54234690
>Current iterations from Pascal are around highend 2015
Current iterations of low end and mainstream are around high-end pascal. I'll be my left testicle that Nvidia is shooting for a Black Friday/Christmas launch time for the Pascal high-end. They're already showcasing the big GP100 pitched for HPCs. Seems like all that is left is to finalizes configurations and hammer out distribution, pricing, and marketing.
>>54234723
>Current iterations of low end and mainstream are around high-end *maxwell
Fixed
>>54234721
>>54234715
Which process is better?
>>54234692
There's a small difference between 16nm TSMC density and 14nm GloFo density, that small difference can't make a 100mm^2 difference, nowhere close.
Diesize is completely dependent on architecture, AMD has a brand new one while Nvidia has a highly modified one, AMD might have optimized for an architecture that won't scale over 400mm2, for example.
They already did that before, and it worked out really well.
>>54234744
For 150W+ GPUs? Wait until June to find out.
Because basing process advantage on a 2W SoC not even using the same process is beyond silly.
It's like comparing 28nm TSMC bulk vs GloFo 28nm SHP
>>54234723
GP100 is in such short supply that Nvidia is only distributing it in their own custom servers for thousands of dollars a piece. OEMs will not get their hands on any GP100 until 2017.
>>54234723
Nah, GP100, aka Big Pascal is only gonna ship in limited quantities, I wager no more than 12000 units in 2016 due to awfully small yields as both the process it's built on is new on top of reaching the reticle limit of the foundry.
Due to these two the price is insane even for HPC GPU.
>>54234750
AMD has also classicly had denser chips even on the same proces, so glofo being ahead in density means they're going to really pack in the hardware.
>>54234839
This is no doubt thanks to their CPU and embeded expertise, no matter how AMD's doing financially, they still have some fantastic engineers.
>>54234823
>>54234839
How many 600nm^2 dies can you fit on a 300mm wafer anyway? Somewhere around 100? How many of those are lost due to random defects and yields? Probably way too much to be acceptable for a company, Intel would rather slit its throat then allow such awful yields.
Nvidia probably rushed GP100 out to prepare the market for NVlink
>>54234216
>GP100 is in such short supply that Nvidia is only distributing it in their own custom servers for thousands of dollars a piece. OEMs will not get their hands on any GP100 until 2017.
The are showcasing binned versions of GP100. That means they are ramping up reproduction for a release with months. The best window is BF/XMAS.
Of course it is going to be in limited quantities in the initial batches but that will anchor price points and generate favorable press (sand another 970 flub).
>>54234908
The binned GP100 is at 300W, obviously Nvidia can't enable the rest of the clusters without balooning the TDP, this obviously points out to bad yields, they will be able to open them by next year, but certainly not now.
It's the same problem as GF100/GF110.
>>54234216
>obviously Nvidia can't enable the rest of the clusters without balooning the TDP
How do you come to this conclusion? IIRC, the binned chip has one of six clusters disabled and pascal takes advantage of the power to perform from maxwell.
How long till we can expect 1080ti release?
I'm in need of a new gpu but knowing that the 1080ti might be another like 5months..
>>54235002
Because more hardware obviously uses more power, duh.
And they've done it before too.
>>54235010
February next year.
Something I've always wondered about binning, a certain amount of shaders/cores is able to be fused off due to defects and not affect the product notably. But what about defects in the memory controller or display controller?
I've never seen , or at least I don't remember a highend GPU's bin having less memory channels.
>>54235033
Wow!
That long, huh?
Boy I should buy a 970 right now shouldn't I?
>>54235100
Those are scrapped, would be too much effort to support more SKUs.
But those are quite rare though, shaders/ALUs make 80% or more of the GPU, so the other parts are designed to be less error prone than the ALUs which have a lot of redundancy.
>>54235033
No shot sherlock. But you seem to believe they will double the TDP to have all six units running. The 970 has one unit cut and the 980' TDP is incrementally higher.
>>54235173
It won't double it, but it will raise it notably and that will affect perf/watt, and server vendors are extremely sensitive to that metric.
Here's the best example of what happens when a chip with bad yields with disabled shaders gets them enabled.
http://www.geeks3d.com/20100810/geforce-gtx-480-512sp-power-consumption-with-furmark/
>>54235201
The 512 core 480 never made it to sale, did it? I can see why.
Mighty impressive how they managed to fix all that with the 580.
>>54235201
That's a very different set of circumstances. It was a rushed process which still had leaks. Pascal has had a full development cycle (more than full thanks to the long period of 28nm). If you're actually functioning under a concept that Pascal is anywhere near the set of circumstances I'm going to point you to everything that happened since then with the 500, 600, and 700 series when it comes to comparing binned to full power versions.
>>54235283
It's the only point of comparison we got, and Nvidia still hasn't officially released a full 60 SM version of GP100(GP110?), which obviously points out that they can't guarantee a defect free GP100 die yet, this is either due to yields or they're keeping the last cluster closed off for a cheap upgrade next year.
In any case, the current binned GP100 will be here for a while because it's what fits the HPC 300W criteria per slot.
>>54234703
Oh baby.
>>54235344
>It's the only point of comparison we got,
I'm sorry? Are you ignoring all the cut down 580 (to be 570s), 680s, 780s and recently 980s that do the exact same thing but hardly makes a difference to the TDP? You to cherry-pick?
>Nvidia still hasn't officially released a full 60 SM version of GP100(GP110?), which obviously points out that they can't guarantee a defect free GP100 die yet
Or they haven't produced enough of them to even start thinking about marketing yet. Or they haven't decided on how far differentiation will be between a full GP100 (Titan GTX let's say) and a cut down 1080Ti.
>>54235389
>I'm sorry? Are you ignoring all the cut down 580 (to be 570s), 680s, 780s and recently 980s that do the exact same thing but hardly makes a difference to the TDP? You to cherry-pick?
These are different things, all of these full dies and bins have been released at the same time ending up in 2 different SKUs, in the case of the 480 512SP and GP100 we know there are unused clusters there but only 1 SKU.
>Or they haven't produced enough of them to even start thinking about marketing yet. Or they haven't decided on how far differentiation will be between a full GP100 (Titan GTX let's say) and a cut down 1080Ti.
If they haven't produced enough full dies but have produced enough binned dies then there's obviously a yield issue afoot, and if there's a large yield issue there's going to be a large power issue as well.
Also, you don't know how good TSMC's process is yet, no 16nm large dies made by TSMC have been sold to anyone to test yet, for all we know both its and Samsung's process is a mess still and next year we'll get 20% higher clocks just due to process improvements.
Nobody here knows this, not you, not me, I can only speculate.
>>54235469
>in the case of the 480 512SP and GP100 we know there are unused clusters there but only 1 SKU.
Weren't the 480 and the 470 two skus just like the Titan and the 780Ti were two skus? Or are you presenting a trivial objection on the technology due to the naming and not the actual chipset itself being used? In the 480/470 case it was Tesla Big or GF100 chipset. In the Titan/780ti it was GK110. The 980ti and Titan X were both from the same GM200 chipset. The difference between the full power cards versus cut cards was negligible in the Titan X to 780ti as both had a 250w TDP. The same is true for Titan X and 980ti. But we should disregard the same chipsets with the same circumstances when there are two different names for the same chipset cards?
>>54235469
>Also, you don't know how good TSMC's process is yet, no 16nm large dies made by TSMC have been sold to anyone to test yet, for all we know both its and Samsung's process is a mess still and next year we'll get 20% higher clocks just due to process improvements.
Aren't all of Samsungs S6, Note 5, and now S7 smartphones produced on the same 16mm finfet and at TSMC?
>>54235627
They're all low power chips, and not all of them are using FF+
>>54234216
> Encode 4K H.265 at 60FPS
How are they doing this exactly? And what will the quality/filesize ratio be like?
>>54235664
The fuck if I know, and I don't care for HEVC encoding, neither for decoding for that fact.
>>54235656
Still, it shows, even circumstantially, that there are some reasons to feel confident in their fabbing.
>>54234296
well, when 980 released it was slightly slower than 780ti, same thing probably will happen with 1080
>>54235700
Would the 980 be consider the flagship of the family? Or even the 980ti? I thought it was the Titan X.
>>54234621
AMd think about dropping CF alltogether when things move to dx12
http://www.youtube.com/watch?v=4qJj1ViyyPY
ARGH waiting for the 490/new fury is hell
How can i make it bros?
I NEED IT NOOOOOOOOOOOAW
>>54235668
I do. I'm frequently encoding videos to reduce their file size.
It's one of the reasons I got my 5960X.
>>54234721
For GPUs AMD will use Samsung as rumors go, it will be very good if they will use glofo also.
>>54235738
Good, CF is pointless bloat.
>>54235674
The only thing that shows is that the process is working.
16nm ffc is incomparable to ff+
2W SoCs are incomparable to 200W GPUs
>>54235664
specialized hardware can do much much more with same number of transistors
forgot the term for that, essentially hardware executes code physically which makes it extremely efficient
there is lecture on computerhistory channel on youtube, they make pretty cool specialized supercomputers that do more for less
>>54235717
yes, until the release of GM200
>>54235852
> specialized hardware can do much much more with same number of transistors
Yes, but the ideal hardware for *encoding* H.264 and H.265 are pretty close to what Intel's and AMD's CPU's already have.
They've already tried using GPU's for encoding before. It hasn't worked that well so far.
>>54235907
no idea, probably they will somehow use what GPU does best - parallel computing
>>54235798
I agree with almost everything you've said.
But those overheating pieces of shit surely have a much higher TDP than just two watts. We must be past five at this point.
>>54235961
2W SDP 5W TDP
During normal workloads those SoCs don't really go much over 900mW outside of bursts, once you push all the cores and rape the cache does the power go up.
>>54234451
I was in your same boat, but decided to go with a 970 for $240 new despite all of the controversy with it. Shit, I might just upgrade to Polaris and then to Vega when the merchandise gets normalized.
So what about the Nvidia side? I'm guessing their 1070 will be 980Ti performance at $400 and their retarded 1080 will be slightly faster at a heftier price. Even if Polaris is $400 it would still be a great buy. Also, wasn't Vega due in between Q42016-Q12017? Shit. I want to upgrade.
On another note, how is AMDGPU nowadays? Last thing I saw was Tonga getting huge boosts, sometimes beating the proprietary driver on Linux. I'd be nice having nice performance on these new cards natively, but still having the option of passthrough if I want to play gaymes.
>>54236154
Pascal will have 1.7% yields, and you won't be able to buy it.
>>54236205
I hope Pascal is another early Fermi, it would be hilarious.
>>54236227
I don't, I want Nvidia to sell the 1080 they've been planning to sell for $500 for $300 due to AMD.
>>54236318
It'll probably be the same prices they used for Maxwell, i.e., $350-$400 for *70 and $530-$550 for *80.
All is meaningless until real benchmarks can be done for both sides. Sadly Nvidia track record is much better than AMD's. They could even do the 3.5 shenanigans without much damage to the bottom line.
Benchmarks or GTFO.
>>54237627
Did you start following Nvidia recently? Fermi was a fucking disaster and Kepler got its shit going way too late with the 780Ti.
