AMD Radeon Pro Duo (Fury X2)

AMD Polaris Gameplay Demo + Fiji Dual GPU 8 GB HBM, 16 TFLOPs
Introducing Radeon Pro Duo


Radeon Pro Duo announced during this AMD GDC Presentation
Capsaicin brought to you by AMD Radeon Graphics
http://www.ustream.tv/channel/gQETjZzLnf9?sf22465712=1

Radeon Pro Duo
For Gamers Who Create and Creators Who Game
http://www.amd.com/en-us/products/graphics/desktop/radeon-pro-duo




More videos:

Spoiler

AMD Radeon Pro Duo Fury X2 Revealed


AMD Demos Polaris Radeon & Vega To Use HBM2 In 2017 | Stunning VR Demo | Radeon Pro Duo | Capsaicin

 

Nice speculation on Guru3d's part

All interesting stuff, up until release on release hardware and optimized drivers we won't really know how it all compares.

I am happy to see AMD working what they have, any way they can. I hope it's priced well and competes with current Nvidia / Pascal releases, and Polaris / Vega release at appropriate timing with Nvidia releases to compete - hitting the mark even before Nvidia releases would be fun too

 

It' all very interesting, the presentation(s) are exciting, I am still watching the Ustream feed, lots of cool stuff
http://www.ustream.tv/channel/gQETjZzLnf9?sf22465712=1



Sulon Q VR/AR headset powered by AMD
http://forum.notebookreview.com/threads/sulon-q-vr-ar-headset-powered-by-amd.789393/

 

Tflops are impressive, I think higher than GTX Titan, I hope they'll have well optimized drivers with it, but where are the laptop cards? Are those cards in OP picture water cooled? Probably running hot.

 

In regards to Polaris... here are some news:
http://arstechnica.com/gadgets/2016/03/amd-gpu-vega-navi-revealed/

The article indicates that it was hinted Polaris will be using HBM1 instead of HBM2... this is contradictory to previous news sources, but then again, was anything in those news confirmed by AMD?
If not, we will have to wait and see for confirmation.

Polaris laptop parts will likely be demo-ed later on, or relatively soon.
Plus, if I'm not mistaken, we SHOULD be getting more powerful Polaris parts as well... when however is a question.

It also seems that 2.5x perf. per watt increase was accurate... meaning that Polaris will likely be able to provide same, if not better performance than Nvidia's Pascal (which is expecting 2x perf. per watt) - at least when you take into account how current Fiji architecture behaves performance-wise, Polaris should be able to provide same/better performance than Pascal.

 

From what I read in other articles it seems to me they'll use combination of HBM and GDDR5 memory. They should have Polaris 10 and 11 for thin and light laptops and for enthusiast: so it seems laptops will get gddr5 and gtx950 class gpu and desktops will get HBM and something around Titan performance, at least at the beginning, which should start shipping around June, if they really want to have full presence for back to school season in August. I also hope they can cut some shaders, downclock it a little and make some high end laptop GPU with HBM memory as well. Anybody knows if Global Foundries making any large scale 14nm chips yet? I know Samsung and Intel is, but they had some issues at the beginning.

 

Pro Duo seems like a waste, with each card limited to 4GB VRAM.

 

The question is do they really need more than 4GB? And I know, some people will say "I have 8Gb and some games use 7GB", but do the games really need it? I don't know the answer, just asking if and why? All I know, for example Win 7 on my laptop takes about 45 GB, but in reality shouldn't use more than 10-15GB, rest is useless garbage, I think it's similar with graphic memory.

 

I don't think games actually need it just yet, unless you are running resolutions at or beyond 4K.
For laptops, HBM1 would probably suffice for top-end mobile gpu Polaris or Vega... though, considering that Vega in the graph exclusively uses HBM2, I think that high end desktop gpu's with Vega will indeed have HBM2... as for laptops... we don't know.
We know nothing about high-end gpu's past Polaris.
Polaris 10 and 11 will be released exactly mid this year, and Polaris 11 will likely give us GTX 970 level performance (which means 15% more performance than 980M but at a much lower power draw).

These Polaris chips I could see as mid-range solutions... whereas high-end mobile chips from AMD will likely be based on Vega and 'might' use HBM2 (whether or not they actually use HBM on laptops remains to be seen as we have nothing on this).

 

If you can bump the vRAM speed up to 6000 or 7000MHz with GDDR5X then I don't see need for HBM in mobile GPU's for the time being. vRAM speed usually isn't the bottleneck. It's either the CPU or GPU performance. The GTX 980 barely budges in benchmark scores if you overclock the vRAM only.

And no, you really don't need 8GB vRAM on a mobile GPU either. I mean, the 980 Ti has 6GB vRAM and doesn't bottleneck it in any way.

 

I'm not too disappointed in the lack or HBM for performance sake. In my opinion the real benefit would have been the power saving for mobile's tpd constraints.

 

The reason the 980 desktop class mxm card is larger is due to the needed space for the vram to clock higher and power circuitry so the limits of mxm are already biting with current GDDR5, increase the power of the core more and it hurts even more.

 

Shouldn't GDDR5X solve this issue though? Lower power, higher bandwidth.

 

That depends on the signaling requirements and how it does on a cramped MXM board and how much of an advantage it offers compared to large increases in core throughput.

 

I think I read that it was built with exactly this in mind, so it wont require redesign and hence the name GDDR5 X.

 

Well, of course . . . Look at Arkham Knight's disastrous texture streaming issue; a problem compounded and epitomised on faster cards with lower vRAM where random stuttering would utterly compromise the experience. Even after the patches, monitoring the GPU resource usage shows that while the game itself does require nearly 5GB of vRAM with everything maximised, it ends up using even more than said limit. Games won't be designed on what we think they ideally should be doing with resources available. With the standards in place, we will have issues relating to inefficiencies -- compounded even further on the PC platform -- and as such, I'd rather err on the side of having more memory to compensate for something, even if in the ideal scenario, such memory wasn't merited.