[WCCF] AMD Radeon R9 390X Pictured

[sm625]

Why is there a curved area of copper?

 

[xthetenth]

Why would apple care if the specs for a discrete video card were leaked?

I can see why AMD would care, but not Apple.

Because they want to unveil its performance in the context of them having it most likely. Rather than the new hotness comes out (oh and Apple has it) it's Apple has the new hotness (and so can you!).

Interesting.

Hard to figure out the die size from that view.
And thats why you are limited to 4 stacks for Fiji and why they can increase capacity in the future by going up in height of exisiting stacks, not multplying stacks

Why can't they go out to put stacks on the side without any? It's pretty much a drop dead certainty that 8 GB would take a new interposer.

 

[maddie]

so roughly 560mm2

 

[maddie]

Interesting.

Hard to figure out the die size from that view.
And thats why you are limited to 4 stacks for Fiji and why they can increase capacity in the future by going up in height of exisiting stacks, not multplying stacks

Because they want to unveil its performance in the context of them having it most likely. Rather than the new hotness comes out (oh and Apple has it) it's Apple has the new hotness (and so can you!).



Why can't they go out to put stacks on the side without any? It's pretty much a drop dead certainty that 8 GB would take a new interposer.

This is one Fury model.

I agree, why can't the other two sides be used? Its not as if a larger interposer is a problem.

 

[Erenhardt]

I see 2 stacks here... Is it that much to ask...just clean this silicon paste damn it!

 

[Udgnim]

I thought HBM stacks on top of each other

 

[twjr]

This is one Fury model.

I agree, why can't the other two sides be used? Its not as if a larger interposer is a problem.

Or even spinning the hbm chips around so the short side is against the gpu. Sure the interposer would have to be lengthened by a few mm but I think you could fit four chips hard up together in the width of the gpu.

 

[dacostafilipe]

I see 2 stacks here... Is it that much to ask...just clean this silicon paste damn it!

There are 2 on each side as far as I can "see" ...

 

[DiogoDX]

I see 2 stacks here... Is it that much to ask...just clean this silicon paste damn it!

4 stacks

 

[Cloudfire777]

This is one Fury model.

I agree, why can't the other two sides be used? Its not as if a larger interposer is a problem.

Try looking at the die shot again. You can see an edge

 

[twjr]

Try looking at the die shot again. You can see an edge

I think they were implying if AMD built a bigger interposer they could used the other edges. But that would leave a lot more empty space (waste) on the interposer. I would assume AMD would want to keep the size of the interposer as small as possible to save cost.

 

[biostud]

What if they could add six layers in each HBM stack?

Most likely it'll be 4Gb and $649 for the hybrid version and $549 for air/cut down version.

 

[ginfest]

I don't think so. If you look at the metal I/O bracket, about 2 of those are the size of the entire PCB. It looks like the card is now 2/3 of the size of an R9 295X2.

I mean it looks significantly smaller in size compared to modern cards.

It's much bigger than Gigabyte's or Galaxy MiniITX 970s:

The size seems closer to the Zotac GTX970, but with a slightly longer PCB where as the Zotac 970's heatsink extends beyond the PCB.

2/3 of R9 295X2 (12.13") => 8.09". I think sub-9 inches seems possible for Fiji XT. Maybe 8.5"?

What does that mean? A flagship card with nearly 300W TDP in what looks like an 8-9" PCB with warrantied AIO CLC that will make it cool and quiet is disappointing to you? We just need performance and price at this point. The design part they absolutely nailed. What did you expect? Another reference blower ~11" card that sounds like a jet engine?

Blowers for flagship cards are not the answer either. Titan Black 250W TDP = 57 dBA at 3100rpm to maintain the same temperatures as an after-market cooler can do at 46-48 dBA at sub-2100 rpm.

If PC gamers embrace AIO CLC for flagship 250W cards, it's going to be possible to make cool and quiet 300W flagships for both AMD and NV, and thus create a CF/SLI setup with most of the heat being exhausted out of the case. For gamers who still want air cooling, there will be open air cooled after-market versions. The blower has really shown to be ineffective for flagship cards like he Titan X once overclocking comes into play. It's not really acceptable to put such a crappy heatsink on premium products anymore as gamers demand superior cooling at high price points.

So we waited what almost 2 years for a bunch of rebrands that won't beat the top performing/selling card (980) in their class and a "super card" with HBM in the over-priced , out of the main stream gamer (Titan) class?
Sorry but that is disappointing, whether the card is 8" or 11.5". These top of the line cards are just for bragging rights, AFAIK they account for less than 5% of the total sold. Unless AMD releases a mainstream card which beats 980 every which way but loose, IMHO it's too little, too late.

 

[Erenhardt]

4 stacks

The two market on the north look like empty space - the surface is not flat.

 

[Azix]

SO they really couldn't find something to wipe that crap off with? Failing.

Everything looks smaller than expected except the HBM being bigger than expected. This chip might be under 500mm^2 The below is a comparison of 5x5 with 22x22 (484mm^2) and 560mm^2. smaller looks more accurate to me.



4GB HBM variant of XT Pro or Hawaii confirmed I guess. Bigger hawaii. Or smaller than expected fiji. Assuming its even an AMD GPU

 

[.vodka]

We do know the size of the HBM chips from this slide:

Let's do some measurements on photoshop with vanishing point. I've measured stuff this way and so far this method hasn't failed me too much, all you need is a good reference point. I took the interposer as the plane, then set the length of one of the HBM chips as a reference point, 7mm. From that, I started measuring the rest.




We know the other side of the HBM chip is 5mm, and the 5.02/5.03mm measurements on both sides of the HBM chip seem to be on spot on where the thermal paste lets us see where the chip looks like it ends. Good, this means we can somewhat trust what's on the picture with the official info in the slides. Now we measure the rest:

GPU: 23.12*25.66mm = 593.25mm².
Interposer: 36.9*27.09mm = 999.62mm²

Taking some margin of error into account, the interposer number is strangely a round one... and the GPU looks to be nearly 600mm²... interestingly near the same size GM200 is. Isn't about 600mm² the supposed limit of what TSMC can manufacture? I don't know where I read that but it rings a bell now.



Just imagine for a second an AMD GPU that is the same size as one of nV's. Throughout those years we've seen RV770/GT200, Cypress/Cayman-GF100/110, Tahiti/GK104, Hawaii/GK110. AMD's GPUs all these rounds got dangerously close to beating or beat nV's best while being quite smaller chips! We've all at some point given a bit of thought to the idea of what would happen if AMD decided to go right up to nV on die size.

If these measurements are anything to go by, Fiji should at the very least make nV uncomfortable. Release it already, AMD!

 

[Vesku]

According to Hynix preview here: http://www.hotchips.org/wp-content/...Bandwidth-Kim-Hynix-Hot Chips HBM 2014 v7.pdf

HBM1 should be 42 mm2

If those HBM stacks are 42mm2 then the GPU die is 500+.

If .Vodka's 7mmx5mm is correct then it still looks around 500mm2, not sure how much to adjust for perspective but eyeballing my bad copy paste I want to guess 24x20 (480mm2) or a bit larger. My second guess is 25x22 (550mm2).

 

[Azix]

everybody's trying something. LoL. I'm getting under 500. The near side of the GPU seems around 24mm the far a lot less. 3.5x7 mm near side 2.5x7mm top side or 4x5mm top side.

 

[Vesku]

everybody's trying something. LoL. I'm getting under 500. The near side of the GPU seems around 24mm the far a lot less. 3.5x7 mm near side 2.5x7mm top side or 4x5mm top side.

Well we know these things are rectangles (Length x Width), the side right next to the HBM (call it Length) does look around 24 to me if the HBM is 7mm the difficult part is getting a good estimate for the Width. .Vodka's perspective analysis is estimating 23 my eyeball is telling me 20-22.

 

[maddie]

I think they were implying if AMD built a bigger interposer they could used the other edges. But that would leave a lot more empty space (waste) on the interposer. I would assume AMD would want to keep the size of the interposer as small as possible to save cost.

Interposer costs = $1 / 100mm2

http://electroiq.com/blog/2012/12/lifting-the-veil-on-silicon-interposer-pricing/

 

[.vodka]

Using vanishing point you only need one correct measure and the plane you've selected. It bases everything off that. If you're accurate enough on selecting these two starting points, you can more or less trust what the program comes up with. I've taken pictures of things at awkward angles like the one we have here, measured this way, then measured the objects themselves with a ruler and the results were more than usable.

The length of the HBM chip and the GPU are thankfully clearly visible in the picture, they have no paste at all. If we believe that slide and use 7mm as our known measurement, that alone is enough for estimating the rest. If we round my measurements:

GPU: 23*25mm = 575mm².
Interposer: 37*27mm = 999mm²


Completely plausible I think. Of course we now have a discrepancy, AMD claims 35mm² (5*7mm) in their slide for a HBM chip, Hynix 42mm² yet they don't give measurements on how they got to that figure in their presentation... We can at least be sure the GPU is over 500mm². It had to be considering the rumored specs (4096sp, etc)

 

[maddie]

According to Hynix the memory stack size is 5.48mm x 7.29mm.

Page 13

http://www.hotchips.org/wp-content/...Bandwidth-Kim-Hynix-Hot Chips HBM 2014 v7.pdf

 

[MrTeal]

http://www.hotchips.org/wp-content/...Bandwidth-Kim-Hynix-Hot Chips HBM 2014 v7.pdf
Page 13.
The 5.48mmx7.29mm is the size of the HBM, including the side mold.

 

[JDG1980]

So we waited what almost 2 years for a bunch of rebrands that won't beat the top performing/selling card (980) in their class and a "super card" with HBM in the over-priced , out of the main stream gamer (Titan) class?
Sorry but that is disappointing, whether the card is 8" or 11.5". These top of the line cards are just for bragging rights, AFAIK they account for less than 5% of the total sold. Unless AMD releases a mainstream card which beats 980 every which way but loose, IMHO it's too little, too late.

I agree that if the rest of the lineup consists of unmodified rebrands of the current TSMC chips, then the 300 series is a fail no matter how Fiji performs. However, if the chips were respun on a lower-leakage Global Foundries process and have the UVD/VCE blocks updated to the newest versions, then the 300 series could be reasonably competitive. AMD could beat the 980 by updating Hawaii with GCN 1.2, upping the shader count to 3072, increasing the default clock to 1100-1200 MHz, and cutting down double-precision support to save die space. If they decide to just release a minor clock bump of the existing Hawaii chip instead, then it'll be a long 18 months while Nvidia continues to eat into their market share and AMD becomes a joke in the enthusiast community.

 

[Azix]

we're talking mm here so might be right. Errors on where the chips end and start included, could be mm difference one way or the other. Wish it was top down cleaned shot.

This person must have been in some kind of rush to take a pic like that, or they are teasing and this is not fiji. Or they don't want to reveal the name of what we call fiji. smh

soon....

Using vanishing point you only need one correct measure and the plane you've selected. It bases everything off that. If you're accurate enough on selecting these two starting points, you can more or less trust what the program comes up with. I've taken pictures of things at awkward angles like the one we have here, measured this way, then measured the objects themselves with a ruler and the results were more than usable.

The length of the HBM chip and the GPU are thankfully clearly visible in the picture, they have no paste at all. If we believe that slide and use 7mm as our known measurement, that alone is enough for estimating the rest. If we round my measurements:

GPU: 23*25mm = 575mm².
Interposer: 37*27mm = 999mm²


Completely plausible I think. Of course we now have a discrepancy, AMD claims 35mm² (5*7mm) in their slide for a HBM chip, Hynix 42mm² yet they don't give measurements on how they got to that figure in their presentation... We can at least be sure the GPU is over 500mm². It had to be considering the rumored specs (4096sp, etc)