[UPDATE] AMD Announces Polaris Architecture - 4th Generation GCN Arriving in Mid 2016

[Notional]

Apparently the first GPU is launching in two months.

http://wccftech.com/amd-polaris/

 

That's the card being demoed after just 2 months from initial production.

[ThreePinkApples]

I like the fact that they didn't say if it was 16nm or 14nm

[don_svetlio]

I like the fact that they didn't say if it was 16nm or 14nm

It's both IIRC

[zMeul]

Apparently the first GPU is launching in two months.

http://wccftech.com/amd-polaris/

that's quite incorrect

WCCFTech making wild assumption .. again - this site should just get banned from being used as a source ever again

[don_svetlio]

that's quite incorrect

WCCFTech making wild assumption .. again - this site should just get banned from being used as a source ever again

[LAwLz]

I like the fact that they didn't say if it was 16nm or 14nm

It's both IIRC

You're right, they will use both.

 

From the Anandtech article:

As for RTG’s FinFET manufacturing plans, the fact that RTG only mentions “FinFET” and not a specific FinFET process (e.g. TSMC 16nm) is intentional. The group has confirmed that they will be utilizing both traditional partner TSMC’s 16nm process and AMD fab spin-off (and Samsung licensee) GlobalFoundries’ 14nm process, making this the first time that AMD’s graphics group has used more than a single fab. To be clear here there’s no expectation that RTG will be dual-sourcing – having both fabs produce the same GPU – but rather the implication is that designs will be split between the two fabs. To that end we know that the small Polaris GPU that RTG previewed will be produced by GlobalFoundries on their 14nm process, meanwhile it remains to be seen how the rest of RTG’s Polaris GPUs will be split between the fabs.

 

So it seems like they might use GlobalFoundries for the lower end cards, and TSMC for the high end cards.

[eLucid]

Apparently the first GPU is launching in two months.

http://wccftech.com/amd-polaris/

 

[LAwLz]

Their assumptions have been accurate lately.

They are going for the shotgun method. They report every single thing and once in a while it turns out to be true.

Even a broken watch is right twice a day...

[mr.squishy]

Every time I see Polaris I think of Super Motherload

[Mahigan]

No, Fermi was just bad design. GCN isn't bad design. It's just dense (HDL-High Density Libraries and all that). While FinFET voltage control is very different, did we all forget what happened to Intel upon moving to FinFET? Heat went up, and overclocking headroom came down. That was true for both the mainstream and HEDT platforms, where HEDT chips were still soldered to their heat spreaders.

Nvidia didn't get it from nowhere. They underwent a paradigm shift in design philosophy. While Koduri hasn't been with AMD for a while, GCN is still based on his previous work. GCN 2.0/4.0 won't be a drastic architectural change other than allowing more TMUs and ROPs as Fiji should have.

No.

Fermi included hardware scheduling components. This made sense, at the time, because Fermi was geared for GPGPU performance.

Kepler ditched most of the hardware based scheduling components as the architecture was geared towards gamers.

Maxwell expanded upon Kepler.

GCN contains hardware based scheduling components as well as ACEs. These all draw power, hence the power usage. This is why GCN's architecture tends to excel at compute workloads.

Polaris, though I can't speak of details, is like GCN but with some shader optimizations. In other words, Polaris can do more with less in terms of compute performance. This opened the door for caching improvements, tessellation improvements as well as a more robust front end. Polaris is different enough from GCN that it can be considered a new architecture. The ALUs are not the same, the Command Processor is not the same, the caching is not the same, the CUs aren't the same, the RBEs aren't the same, the memory controller is brand new etc etc etc.

This test was conducted with a frame rate cap of 60FPS. The figures, AMD obtained, are entirely within the realm of possibility. We're looking at up to 2.5 times the performance per watt of Fiji with Polaris.

Pascal, given that it is Maxwell version 2, should retain its excellent performance per watt figures but will likely be beaten in terms of pure compute performance by Greenland.

My take? Greenland overtakes Pascal.

[Dabombinable]

No.

Fermi included hardware scheduling components. This made sense, at the time, because Fermi was geared for GPGPU performance.

Kepler ditched most of the hardware based scheduling components as the architecture was geared towards gamers.

Maxwell expanded upon Kepler.

GCN contains hardware based scheduling components as well as ACEs. These all draw power, hence the power usage. This is why GCN's architecture tends to excel at compute workloads.

Polaris, though I can't speak of details, is like GCN but with some shader optimizations. In other words, Polaris can do more with less in terms of compute performance. This opened the door for caching improvements, tessellation improvements as well as a more robust front end. Polaris is different enough from GCN that it can be considered a new architecture. The ALUs are not the same, the Command Processor is not the same, the caching is not the same, the CUs aren't the same, the RBEs aren't the same, the memory controller is brand new etc etc etc.

This test was conducted with a frame rate cap of 60FPS. The figures, AMD obtained, are entirely within the realm of possibility. We're looking at up to 2.5 times the performance per watt of Fiji with Polaris.

Pascal, given that it is Maxwell version 2, should retain its excellent performance per watt figures but will likely be beaten in terms of pure compute performance by Greenland.

My take? Greenland overtakes Pascal.

Considering Maxwell already loses out in compute performance to GCN, I agree.

[TheRandomness]

So I can assume this new architecture won't be good because they've been alternating between good and bad every generation-

tahiti- good

hawaii- bad

fiji- good

polaris- bad?

Hawaii was/is great without the stock cooler

[patrickjp93]

No.

Fermi included hardware scheduling components. This made sense, at the time, because Fermi was geared for GPGPU performance.

Kepler ditched most of the hardware based scheduling components as the architecture was geared towards gamers.

Maxwell expanded upon Kepler.

GCN contains hardware based scheduling components as well as ACEs. These all draw power, hence the power usage. This is why GCN's architecture tends to excel at compute workloads.

Polaris, though I can't speak of details, is like GCN but with some shader optimizations. In other words, Polaris can do more with less in terms of compute performance. This opened the door for caching improvements, tessellation improvements as well as a more robust front end. Polaris is different enough from GCN that it can be considered a new architecture. The ALUs are not the same, the Command Processor is not the same, the caching is not the same, the CUs aren't the same, the RBEs aren't the same, the memory controller is brand new etc etc etc.

This test was conducted with a frame rate cap of 60FPS. The figures, AMD obtained, are entirely within the realm of possibility. We're looking at up to 2.5 times the performance per watt of Fiji with Polaris.

Pascal, given that it is Maxwell version 2, should retain its excellent performance per watt figures but will likely be beaten in terms of pure compute performance by Greenland.

My take? Greenland overtakes Pascal.

 

You don't need hardware scheduling for good GPGPU performance. The Kepler Teslas prove as much being the top used accelerator in the world for compute tasks. The only thing hardware scheduling is particularly useful for is Asynchronous Compute/Shading pipelines.

 

GCN doesn't excel at compute workloads. They're outclassed by a Knight's Corner Xeon Phi in Linpack and HPC applications despite the Xeon Phi having 1/3 the theoretical performance cap. Either bad drivers or bad design, but they most certainly are not compute monsters as everyone seems to think.

 

Only theoretically. AMD has a lot of catching up to do to make their compute competitive.

[shawno]

You don't need hardware scheduling for good GPGPU performance. The Kepler Teslas prove as much being the top used accelerator in the world for compute tasks. The only thing hardware scheduling is particularly useful for is Asynchronous Compute/Shading pipelines.

GCN doesn't excel at compute workloads. They're outclassed by a Knight's Corner Xeon Phi in Linpack and HPC applications despite the Xeon Phi having 1/3 the theoretical performance cap. Either bad drivers or bad design, but they most certainly are not compute monsters as everyone seems to think.

Only theoretically. AMD has a lot of catching up to do to make their compute competitive.

Where do i buy a xeon phi graphics card to run my games?

[Fetzie]

Where do i buy a xeon phi graphics card to run my games?

 

You can't, they are co-processors for certain types of server motherboard.

[Clear17Mud]

LOL I found these post hilarious... I love how it got so off topic from the video, and also I would guess that if anything the videos are true but the amd system is getting measured at a lower power usage time, and the nvidea system is at a higher load time, but also the cpu is capped at 80%.

another thing @patrickjp93 saying that nvidea crushes or AMD crushes(not you but others) over the past few years they all performed relatively good for their prices, but amd had a lot more price to performance cards in the mid range, and the R9 290 was basically the 980ti in 980ti vs Titain compared to R9 290 vs R9 290x. because they releases a full core, but at a lower clock. also a r9 fury x may not compare to a 980ti very well but a non x fury is $100+ less. 

 

Also @patrickjp93 you are a very opinionated person... you made me laugh the most... its like you have a tin hat for amd, and then had nvidea and intels mistakes surgically removed from your mind... but I guess to each their own, and I'm prepared for your flaming...

[patrickjp93]

Where do i buy a xeon phi graphics card to run my games?

 

They aren't GPUs, not unless you use a very efficient software-based rendering engine, and even then you need to send the image data through the motherboard graphics output.

 

Xeon Phi are built purely as compute accelerators.

[patrickjp93]

LOL I found these post hilarious... I love how it got so off topic from the video, and also I would guess that if anything the videos are true but the amd system is getting measured at a lower power usage time, and the nvidea system is at a higher load time, but also the cpu is capped at 80%.

another thing @patrickjp93 saying that nvidea crushes or AMD crushes(not you but others) over the past few years they all performed relatively good for their prices, but amd had a lot more price to performance cards in the mid range, and the R9 290 was basically the 980ti in 980ti vs Titain compared to R9 290 vs R9 290x. because they releases a full core, but at a lower clock. also a r9 fury x may not compare to a 980ti very well but a non x fury is $100+ less. 

 

Also @patrickjp93 you are a very opinionated person... you made me laugh the most... its like you have a tin hat for amd, and then had nvidea and intels mistakes surgically removed from your mind... but I guess to each their own, and I'm prepared for your flaming...

I'm not tin foil hat on AMD or forgetful of Nvidia's mistakes. I always base my arguments on what is, not what was.

 

Intel buying off OEMs? Otellini and Huddy stepped down just shy of a decade ago. Same issue with the Pentium 4 (iirc) floating point problem: the perpetrators are long gone.

 

Nvidia melting GPUs? 7 years old now? I don't consider the 970 issues to be either problematic nor maliciously intended. Clerical errors will happen to any company EVENTUALLY. I'd say the same if it happened to Intel or AMD. Gameworks clearly isn't made to cripple AMD since there are a couple GW titles AMD is better at.

 

AMD's complete hype/lies about Barcelona are also that old. Bulldozer is much more recent in history, and the engineers responsible for it still work at AMD. 290/X reference heat problems are also still current. AMD's driver problems are more current, with Crimson having fried a few GPUs as well. AMD didn't deliver on open-source Mantle as promised. AMD has yet to deliver on GPUOpen or fully O-S graphics drivers for Windows or Linux. We'll have to wait and see on that end.

 

No flaming required, just facts.

[shawno]

They aren't GPUs, not unless you use a very efficient software-based rendering engine, and even then you need to send the image data through the motherboard graphics output.

Xeon Phi are built purely as compute accelerators.

Forgot to add /s

[TechGod]

mate, you're saying that both testing are capped - that means the GTX950 would draw a significant amount less power than it's actually rated for

keeping in mind that both systems minus videocards would draw equal amounts of power, that would make system 1 draw power from where!? a parallel universe!?

let's say the GTX950 would draw on it's own, capped, ~50W - that would mean the rest of the system would draw 100W

and this is where all bells, whistles and horns would ring because system1, with video card, draws only 85W - HELLO!

as I said, AMD's testing is bullcrap!

system1 is running with all power caps in place, disabled core, disabled lanes

system2 is running with no caps

How do you know AMD disabled things for one system and not for the other? You'd believe it if it was Nvidia you dumbass fanboy.

[Mahigan]

You don't need hardware scheduling for good GPGPU performance. The Kepler Teslas prove as much being the top used accelerator in the world for compute tasks. The only thing hardware scheduling is particularly useful for is Asynchronous Compute/Shading pipelines.

GCN doesn't excel at compute workloads. They're outclassed by a Knight's Corner Xeon Phi in Linpack and HPC applications despite the Xeon Phi having 1/3 the theoretical performance cap. Either bad drivers or bad design, but they most certainly are not compute monsters as everyone seems to think.

Only theoretically. AMD has a lot of catching up to do to make their compute competitive.

I believe I've mentioned this to you on several instances. Software.

AMD lacks proper software optimization tools. AMD rely on OpenCL.

Knights Landing/Xeon Phi can also be powered by OpenCL but its main software stack are in

OpenMP, MPI, C, C++ and Fortran which are all more robust programming languages.

CUDA is still more robust than OpenCL as well.

In terms of the hardware, though, GCN is a monster compared to Kepler and Maxwell in terms of GPGPU. Heck Hawaii, R9 290x, outperforms the Kepler based Titan Black under OpenCL and DirectCompute.

GCN is a compute oriented architecture but which made a few compromises in order to handle Gaming Tasks as well. One such compromise, compared to Tahiti, is Hawaii and Fiji's FP64 performance.

Polaris builds upon GCN. When Greenland arrives, you will see what I mean. Larger registers and cache should mean better FP64 performance.

What AMD needs, in the realm of HTPCs, is more robust OpenCL support. This will take time and will arrive as OpenCL matures.

http://us.hardware.info/reviews/5231/8/nvidia-geforce-gtx-titan-black-4-way-sli-review-gpgpu-benchmarks