AMD: Pirate Islands (R* 3** series) Speculation/Rumor Thread

[silent_guy]

Margins schmargins. You have the AMD execution woes thread to talk business; this is the GPU board thread.

The margins are a direct consequence of the architecture.

That said, even if their margins take a (likely minor; these water coolers aren't exactly made of unobtanium) hit, having a premium product with cutting edge performance is still beneficial, as it builds their presence in the high performance market and builds goodwill with hardware enthusiasts; both of which is sorely lacking at the moment.

I agree.

What does it matter how they obtain their efficiency improvement as long as they do obtain it?

If you aren't interested in GPU architecture and how performance is achieved, WTF are you doing in a subforum called '3D architecture & chips' ?

Even so, your concern would seem to be unfounded, as HBM alone could not possibly represent a full 50% increase in perf/watts. That would mean GDDR bus and chips in Hawaii ate up more than half the board power, which is simply absurd.

Now we're having a discussion.

We'll have to see what has happened with perf/W. But it's not only that of course.

I see HBM as a fantastic invention that will eventually give a much needed boost to GPU. I think AMD deserves tons of credit for pulling it off, at least the part where they get HBM to work at all.

But what we're currently seeing is that despite a 80% increase in BW compared to its competing chip, it's barely getting ahead. It's as if AMD bolted a fuel tank with a humongous intake onto an engine that doesn't know what to do with it. IOW that all this marvelous work on HBM is only a bandage on a pretty inefficient core architecture.

Maybe it's worth doing for marketing brownie points, just like the water cooler, but it means that Fiji didn't manage to solve the core issues with GCN, and that they won't have the HBM advantage for the next generation. They still have a lot of work to do.

 

[Babel-17]

Except, power draw isn't directly comparable to benchmark numbers, so I don't really see your point.


There's a lot of negativity radiating from you in this thread. What is it they're "hiding", you think? They announced 1.5x perf/watt improvement vs. hawaii. Other figures for power useage is also circulating in this thread, so I have to assume those also come from AMD. Sounds like you got an axe of some sorts to grind, for reasons I can't quite figure out.


Why would you need specific BIOS support for new video cards? All video cards use the same BIOS compatibility nonsense hailing right back to the early 1980s, to BIOS, even the latest GPUs are just a dumb text display. So you should be totally fine with any video card you like, provided Dell provided you with the appropriate auxiliary PCIe power connectors, which isn't a surefire thing. (Saying this from experience; had an XPS box with a 1000W PSU that only offered 6-pin power connectors. BLAH.)

Well, it wasn't the PSU's, a lot of the people freaking out had bought powerful, and high quality, PSU's just for their new shiny cards. And fwiw, the stock Dell PSU had two six pin connectors and something like 32 amps on the 12v rail (a 460 watt PSU, maybe a Delta). Anyway, nVidia did come in and did work with Dell on the bios.
I now see I was off on the model that got the aid, and the nVidia series that was causing the problem. It was the XPS 8700, and using the 9xxx series cards. IIRC it got the help because Dell was still selling a few when the 9xxx cards were released. The agony of the 8300 series is another tale.

http://en.community.dell.com/support-forums/desktop/f/3515/t/19601491?pi21953=1
Feast your eyes on that outer circle of hell. Lol, I know, rich peoples problems, but the screams are real.

But in a nutshell, here:

Posted by DELL-Chris M
on 29 Sep 2014 11:01 AM

UPDATE 12/15/14 = The A10 bios is here and on the XPS 8700 page here.

* Download/Save the file to the windows desktop
* Before running A10, I would install the prior Bios in order
* Disconnect any external hard drives, keys, printer, scanner, etc. You only want the USB mouse, USB keyboard, LAN, monitor connected
* Run the executable from Windows
* Allow the system to complete the installation and restart automatically. Do not force a restart or power off
===============================

NEW thread started to discuss XPS 8700 Nvidia GTX video card upgrades

Place holder for new thread to discuss XPS 8700 AMD video card upgrades
===============================

We and Nvidia are looking into the issue, but understand, we only "test/validate" those video cards in my list. Our warranty does not cover non-OEM video cards NOT on our list. Ultimately, it will be up to the XPS 8700 Team and our Legal Team to decide if Dell is going to allocate resources to fix an issue with non-validated, non-tested 3rd party hardware. All I can do is funnel the data to them. Below are the Dell validated/tested OEM video cards for the XPS 8700 and our 460w power supply =

8MXMJ Nvidia GeForce GTX750Ti P123N, 2GB, DVI-I/DVI-D/mHDMI
TC2P0 Nvidia GeForce GTX745 M302N, 4GB, DVI-I/DP/HDMI
9YJWT Nvidia GeForce GT720 M211N, 1GB, VGA/DVI-D/HDMI
FPDH3 Nvidia GeForce GTX660 D15P2-40 V2, 1.5GB, DVI-I/DVI-D/DP/HDMI, (Win 8)
2CHCY Nvidia GeForce GTX660 D14P2-40, 1.5GB, DVI-I/DVI-D/DP/HDMI, (Win 8)
8C3R5 Nvidia GeForce GTX650Ti P122N, 1GB, DVI-I/DVI-D/mHDMI
9168H Nvidia GeForce GTX645 OHGA4/M301N, 1GB, DVI-I/DP/HDMI
R5H2D Nvidia GeForce GT635 M210N, 1GB, VGA/DVI-D/HDMI

9KYFK AMD Radeon R9 270 RD14D-P1-70, 2GB, DVI-I/DVI-D/DP/HDMI
WNMHJ AMD Radeon HD8870, 2GB, DVI-I/mDP/mDP/HDMI
51NCR AMD Radeon HD7570 M209A, 1GB, DVI-I/HDMI (Win 8)



(Liaison for Desktops, Alienware, Monitors)#IWork4Dell
Manuals and Drivers
Technical Support via Chat
Technical Support via Facebook

And ironically, I'm an 8300 series owner using an older BIOS, and with a 9700.

Edit: Hmm, maybe it was because the 8700 had an early version of UEFI, and in conjunction with the way Windows 8 checks over a system when it boots up, the video card wasn't passing the smell test. Yeah, I think it was something technical, and this is just an attempt to point you in the right direction as to what was what. IIRC it got discussed within the labyrinth of that thread. It must have been excruciating for those Dell owners. I went through something similar getting eSATA support working for my non-RAID XPS 420. But it was resolved.
Now that we know things worked out that thread can be said to have some amusing features. You know, comedy is tragedy from a safe distance.
The screams of outrage as consumers ran into the brick wall of Dell tech support ... Really, an outer circle of geek hell.

 

[Alexko]

Any idea when the firepro versions of this will be available? Looking for something that doesnt have gimped double precision.

Sadly, the low DP rate seems to be a hardware limitation, or so it says here:
http://www.hardware.fr/news/14263/amd-radeon-r9-fury-x-gpu-fiji-details.html

 

[silent_guy]

Sadly, the low DP rate seems to be a hardware limitation, or so it says here:
http://www.hardware.fr/news/14263/amd-radeon-r9-fury-x-gpu-fiji-details.html

Ultimately, Nvidia and AMD had to fight the same limitation in terms of die size. Something has to give...

 

[Grall]

It was the XPS 8700, and using the 9xxx series cards. IIRC it got the help because Dell was still selling a few when the 9xxx cards were released. The agony of the 8300 series is another tale.

From what that support guy wrote in your quote, it sounds a lot like Dell actually put some kind of whitelist in their BIOS (UEFI?), listing cards it has been tested/validated with, and possibly even deliberately stopping POST with unsupported cards...?

You really don't need specific BIOS/UEFI support for new hardware, assuming your video card and PC conforms to mutual form factor and electrical specs requirements... If you did, the whole PC build-it-yourself scheme would have come crashing down decades ago, and no (advanced) video card vendors could have survived, due to a lack of systems to stick new models in... We'd all probably still be sitting with plain-jane CPU-driven framebuffer video cards. (The horror!)

 

[Alexko]

Ultimately, Nvidia and AMD had to fight the same limitation in terms of die size. Something has to give...

Yes, that seems to be the primary motivation. It probably saves a few watts too.

 

[Babel-17]

From what that support guy wrote in your quote, it sounds a lot like Dell actually put some kind of whitelist in their BIOS (UEFI?), listing cards it has been tested/validated with, and possibly even deliberately stopping POST with unsupported cards...?

You really don't need specific BIOS/UEFI support for new hardware, assuming your video card and PC conforms to mutual form factor and electrical specs requirements... If you did, the whole PC build-it-yourself scheme would have come crashing down decades ago, and no (advanced) video card vendors could have survived, due to a lack of systems to stick new models in... We'd all probably still be sitting with plain-jane CPU-driven framebuffer video cards. (The horror!)

It's puzzling, it's not like Dell doesn't have to do this for every new model. And for it actually taking nVidia people to get involved suggests that there was something tricky to this. Maybe Dell was also worried about legal liability and their lawyers would only accept solutions that kept them 100% liability free? Just spitballing here, and OT, but I appreciate your response and didn't want to ignore it. I guess we'll have to chalk this up to ....

 

[lanek]

Is disabling the Uefi check will not be enough in this case ?

 

[Razor1]

http://www.legitreviews.com/amd-fiji-arrives-radeon-r9-fury-x-details_166515#OpQZsfoAmLt9ajjt.99

AMD feels that the memory technology is too new and there is more than enough bandwidth, so they are locking down the ability to overclock the memory in AMD Overdrive.

So the chip won't be bandwidth bottlenecked? That should be interesting.

 

[madyasiwi]

Not to my knowledge...

This tweet from sometimes ago could give a hint, crossfire setup:
https://twitter.com/draginol/status/567428591630426112

 

[lanek]

http://www.legitreviews.com/amd-fiji-arrives-radeon-r9-fury-x-details_166515#OpQZsfoAmLt9ajjt.99



So the chip won't be bandwidth bottlenecked? That should be interesting.

How you will 512 GB/s be bottlenecked .. i was allready nearly sure that it will not be overclockable, because due to the nature of HBM, it is really complicate on the voltage side. on the power electricity side, it is extremely complicated. It is like nothing we have seen before, and playing with clock speed and voltage on HBM is really really risky.

 

[Razor1]

How you will 512 GB/s be bottlenecked .. i was allready nearly sure that it will not be overclockable, because due to the nature of HBM, it is really complicate on the voltage side. on the power electricity side, it is extremely complicated.

Well if they are making this card for overclocking and don't let people over clock the ram what do you think will happen?

 

[Malo]

Got any links?

edit: ugh nevermind. They are Grenada (Hawaii) card reviews. Nothing for Fiji yet of course. Got my cards mixed up.

 

[Ethatron]

But what we're currently seeing is that despite a 80% increase in BW compared to its competing chip, it's barely getting ahead. It's as if AMD bolted a fuel tank with a humongous intake onto an engine that doesn't know what to do with it. IOW that all this marvelous work on HBM is only a bandage on a pretty inefficient core architecture.

You can make a chip for the current workload characteristics, or worst real world implementations. Or you make one for a different vision, for which you have to change the algorithms, leave the current stuff.

Pick your poison.

 

[silent_guy]

Well if they are making this card for overclocking and don't let people over clock the ram what do you think will happen?

If it's completely over-provisioned in terms of BW, why would they allow it?
That said: it's very disappointing that they don't allow to underclock it, so we can't see how excessive the BW really is.

 

[silent_guy]

You can make a chip for the current workload characteristics, or worst real world implementations. Or you make one for a different vision, for which you have to change the algorithms, leave the current stuff.

Pick your poison.

As if that were the 2 only options...

 

[Razor1]

If it's completely over-provisioned in terms of BW, why would they allow it?
That said: it's very disappointing that they don't allow to underclock it, so we can't see how excessive the BW really is.

well going by numbers from the r290x, its got ~60% increased ALU throughput and ~60% increased bandwidth, seems to be it should be fairly equal to what we see now when it comes to bottlenecks, but of course as you said before its hard to compare between the generations this time.

 

[Ethatron]

As if that were the 2 only options...

It's only a suggestion that it's not the task of a new GPU to accelerate the stuff from yesteryear.

 

[silent_guy]

i was allready nearly sure that it will not be overclockable, because due to the nature of HBM, it is really complicate on the voltage side. on the power electricity side, it is extremely complicated. It is like nothing we have seen before, and playing with clock speed and voltage on HBM is really really risky.

This is all new to me. Tell us more!

 

[Razor1]

http://cdn.videocardz.com/1/2015/06/AMD-Radeon-R9-Fury-X-Overclock-performance.png

more benches

10% overclock on the core is giving ~5% performance increase according to this, so memory overclocking please.....