[WCCF] AMD Radeon R9 390X Pictured

[.vodka]

This one is about perfect. Well defined edges, finally! And it is clear here the shiny parts are the interposer. This is because we know the HBM chips are rectangular (7.29*5.48mm) while the shiny parts are more like squares. No exposed silicon means these HBM chips have the mold (packaging) as described in Hynix's presentation, so these should be the actual measurements for the HBM chips.

This time the plane can be set more clearly and is less extreme, and the clearest base measure is 5.48mm on one side of the HBM chips. Much less guessing now, means more accurate results. And as always photoshop's vanishing point takes into account perspective. I've sharpened up the picture to make the edges even clearer. Let's see:



GPU: 24.01*27.72mm = 665.55mm²
Interposer: 41.48*29.83mm = 1231.38mm²

Rounding as margin of error:

GPU: 24*28mm = 672mm²
Interposer: 41*30mm = 1230mm²


My measurements on WCCFtech's angled picture threw similar estimates. We can all be pretty sure this thing is at least over 600mm² accounting for some big margin of error (a mm or 2 changes everything).

--------------------------------------------

With that die size (650-700mm²) nV should be scared, to put it lightly. For the past years they've competed against AMD with far bigger chips, not beating AMD's competition conclusively for their added size, usually getting tied after some time and lately losing... A quick google search throws these numbers:


RV770: 256mm²
GT200: 576mm²

Cypress: 336mm²
GF100: 520mm²

Cayman: 389mm²
GF110: 520mm²

Hawaii: 438mm²
GK110: 551mm²

And now...
Fiji: 600-700mm²
GM200: 600mm²


I don't count G80-G92/R600-RV670 because we all know how that went. But the architecture there turned out to be something interesting after some refinement. Here we have a proven one, GCN, which can only get better, and we've seen that with Tonga. I also don't count Maxwell here because Hawaii is Kepler's competition.

This ridiculously big thing right here is what GM200 is up against... I don't think it'll be pretty for nV this time. It's a lot of time until Pascal+HBM2 is here, AMD will probably keep the crown up to that point with such a big GPU plus the new memory technology. I can only wonder how much nV would have been embarrased that generation if RV770 was at least as big as the 5870's Cypress using brand new GDDR5 as it did back then. Thankfully this time AMD isn't doing the small die approach with HBM.

The question here is, can AMD make up for all these months of Maxwell taking their market share, to at least revert the situation back to when Maxwell launched? Hell, I wonder if Fiji will be enough cut through the apple-like reality distortion field nV has created on their users, let alone the loyal ones... I can't wait for E3 to come!

 

[Azix]

That's still pretty inaccurate thanks to the crap resolution.

 

[Silverforce11]

This one is about perfect. Well defined edges, finally! And it is clear here the shiny parts are the interposer. This is because we know the HBM chips are rectangular (7.29*5.48mm) while the shiny parts are more like squares. No exposed silicon means these HBM chips have the mold (packaging) as described in Hynix's presentation, so these should be the actual measurements for the HBM chips.

This time the plane can be set more clearly and is less extreme, and the clearest base measure is 5.48mm on one side of the HBM chips. Much less guessing now, means more accurate results. And as always photoshop's vanishing point takes into account perspective. I've sharpened up the picture to make the edges even clearer. Let's see:



GPU: 24.01*27.72mm = 665.55mm²
Interposer: 41.48*29.83mm = 1231.38mm²

Rounding as margin of error:

GPU: 24*28mm = 672mm²
Interposer: 41*30mm = 1230mm²


My measurements on WCCFtech's angled picture threw similar estimates. We can all be pretty sure this thing is at least over 600mm² accounting for some big margin of error (a mm or 2 changes everything).

--------------------------------------------

With that die size (650-700mm²) nV should be scared, to put it lightly. For the past years they've competed against AMD with far bigger chips, not beating AMD's competition conclusively for their added size, usually getting tied after some time and lately losing... A quick google search throws these numbers:


RV770: 256mm²
GT200: 576mm²

Cypress: 336mm²
GF100: 520mm²

Cayman: 389mm²
GF110: 520mm²

Hawaii: 438mm²
GK110: 551mm²

And now...
Fiji: 600-700mm²
GM200: 600mm²


I don't count G80-G92/R600-RV670 because we all know how that went. But the architecture there turned out to be something interesting after some refinement. Here we have a proven one, GCN, which can only get better, and we've seen that with Tonga. I also don't count Maxwell here because Hawaii is Kepler's competition.

This ridiculously big thing right here is what GM200 is up against... I don't think it'll be pretty for nV this time. It's a lot of time until Pascal+HBM2 is here, AMD will probably keep the crown up to that point with such a big GPU plus the new memory technology. I can only wonder how much nV would have been embarrased that generation if RV770 was at least as big as the 5870's Cypress using brand new GDDR5 as it did back then. Thankfully this time AMD isn't doing the small die approach with HBM.

The question here is, can AMD make up for all these months of Maxwell taking their market share, to at least revert the situation back to when Maxwell launched? Hell, I wonder if Fiji will be enough cut through the apple-like reality distortion field nV has created on their users, let alone the loyal ones... I can't wait for E3 to come!

Even if Fiji is a lot faster, top tier SKUs don't move market share much. For NV, the damage was done by the 960 and 970. AMD needs competitive parts in those segments if they wish to claw back marketshare and have a nice revenue.

I have hope the "remake" of Tonga/Hawaii on GloFo will allow them to be competitive on perf/w & perf/$ against 960/970/980.

 

[eRacer]

Certainly doesn't look like a heatspreader to me.

Yep, no heatspreader. It would be shiny rather than black and cover the GPU core and HBM. That's what I get for posting at 1AM instead of sleeping.

 

[.vodka]

Yes, but it is clearly something above 600mm², the crappy resolution in all the pictures we have is enough for that lower limit. Such a die size still validates my opinion and prediction.

That ridiculously big GPU right there is exactly why the 980Ti even exists, providing the same performance as the X at $350 less. At least nV will get some sales from here to when AMD releases their new card in three weeks.

Of course, we can be pretty certain they'll have the performance crown. But pricing will make or break it... If AMD gets the pricing well, they'll regain their precious market share. I hope they do, this 80/20 situation as it is right now is unsustainable for the ecosystem and gaming community going forward.


And as silverforce says, I also hope the rest of the lineup had better put up a good fight...

 

[RaulF]

Even if Fiji is a lot faster, top tier SKUs don't move market share much. For NV, the damage was done by the 960 and 970. AMD needs competitive parts in those segments if they wish to claw back marketshare and have a nice revenue.

I have hope the "remake" of Tonga/Hawaii on GloFo will allow them to be competitive on perf/w & perf/$ against 960/970/980.

Well if AMD did some tweaks to Hawaii(added processing units, added some tonga stuff to it) and with the suppose better process at GoFlo, it could be good competition for 960/970. Time will tell.

 

[flopper]

The question here is, can AMD make up for all these months of Maxwell taking their market share, to at least revert the situation back to when Maxwell launched? Hell, I wonder if Fiji will be enough cut through the apple-like reality distortion field nV has created on their users, let alone the loyal ones... I can't wait for E3 to come!

The Fury pro be the card to own.
Unless they also add a third one into the mix.
Yea entusiast buy the fastest no matter the brand.

seems they gone all out this time.
fastest, new technology why buy old slower technology?

 

[thehotsung8701A]

AMD should make an even better GPU call it the FURRY MAX PRO XT. They haven't use XT since Ati day though. Since it call furry, release it with a better TressFX technology as well.

Joking aside, just saw on a tweet that all new AMD GPU will come with Dirt Rally?

 

[RussianSensation]

I wonder what certain posters from Team Green will say if Fiji XT is 5% slower than the Titan X but has 10-15X the double precision compute performance. After-all, one of the major justifications for the OG Titan's $1k price tag was compute since it made it a semi-pro workstation card. I guess if Fiji offers 95% of the gaming performance of the Titan X at 4K and offers a lot more compute, the dual-purpose functionality of the card will NO longer matter and compute advantages will be dusted under the rug.......until Pascal brings back FP64 in full force for NV and then we'll hear again how Compute in a gaming card deserves premiums since it makes it a semi-pro card.

Also, certain posters already should be prepared to be eating crow:

> This will be AMD's largest GPU die of all time - crow served to those who said AMD won't make a 500mm2+ die chip.

> Perf/watt will be greatly improved from GCN 1.0/1.1 based on the HUGE die size alone and 300W TDP - crow served to those who said AMD will not be able to improve perf/watt on any next gen 28nm cards. Per Silverforce11's quote: "However, AMD's Graphics CEO Joe Macri told Golem.de that Fiji graphics card should not require more energy than the Radeon R9 290X." :whiste:

 

[Alatar]

I wonder what certain posters from Team Green will say if Fiji XT is 5% slower than the Titan X but has 10-15X the double precision compute performance. After-all, one of the major justifications for the OG Titan's $1k price tag was compute since it made it a semi-pro workstation card. I guess if Fiji offers 95% of the gaming performance of the Titan X at 4K and offers a lot more compute, the dual-purpose functionality of the card will NO longer matter and compute advantages will be dusted under the rug.......until Pascal brings back FP64 in full force for NV and then we'll hear again how Compute in a gaming card deserves premiums since it makes it a semi-pro card.

AMD has artificially crippled FP64 on their consumer cards since hawaii so I don't see how this would matter at all for a consumer.

If you want a consumer card that's good at FP64 you buy tahiti at the lower end or Titan/Titan black at the higher end. Fiji is unlikely to change this.

Also, certain posters already should be prepared to be eating crow:

> This will be AMD's largest GPU die of all time - crow served to those who said AMD won't make a 500mm2+ die chip.

Pretty much no one has said this since the synapse leaks showing 500mm^2+ nearly a year ago

> Perf/watt will be greatly improved from GCN 1.0/1.1 based on the HUGE die size alone and 300W TDP - crow served to those who said AMD will not be able to improve perf/watt on any next gen 28nm cards. Per Silverforce11's quote: "However, AMD's Graphics CEO Joe Macri told Golem.de that Fiji graphics card should not require more energy than the Radeon R9 290X." :whiste:

Power consumption and efficiency remains to be seen. However it's going to be almost impossible to claim that the GCN in Fiji has better or worse perf/watt than any other AMD GCN part because HBM is going to completely ruin any comparisons between the architectural revisions. Fiji itself will almost certainly have better perf/watt but the chances of anyone being able to say for certain that it's efficiency for GCN itself are low.


Either way the real deal when it comes to eating crow are:

-8GB
-Titan X crushing performance for Fiji
-Multiple HBM based GPUs
-Completely new GPU lineup
-GCN architectural improvements over tonga
-new GPUs and competitiveness for mGPU market
-GCN 1.2 across the product stack / GCN 1.2 for hawaii
-glofo

 

[Silverforce11]

Power consumption and efficiency remains to be seen. However it's going to be almost impossible to claim that the GCN in Fiji has better or worse perf/watt than any other AMD GCN part because HBM is going to completely ruin any comparisons between the architectural revisions. Fiji itself will almost certainly have better perf/watt but the chances of anyone being able to say for certain that it's efficiency for GCN itself are low.

Either way the real deal when it comes to eating crow are:

-8GB
-Titan X crushing performance for Fiji
-Multiple HBM based GPUs
-Completely new GPU lineup
-GCN architectural improvements over tonga
-new GPUs and competitiveness for mGPU market
-GCN 1.2 across the product stack / GCN 1.2 for hawaii
-glofo

AMD seems to think it won't use more power than R290X, if the full German quote is taken and translated, it continues, saying that there's a very good chance it will use LESS power than R290X. To put that in context, non-reference R290X uses ~230-250W gaming load. Titan X uses ~235W gaming load.

Pretty beastly efficiency gains.

The rest of the line-up, we'll have to see if its GloFo or TSMC, if its GloFo, chances are very high that its more than competitive against 960,970 and 980. If its TSMC, no way.

GloFo has confirmed Fiji at least is made by it ("2.5D stacking & HBM massive GPU"). I think odds are very high Grenada & Trinidad are also on GloFo due to their big $1B wafer agreement.

 

[dacostafilipe]

Hardware.fr confirms 4GB limit on Fiji, clamshell not possible width HBM1 but will be with HBM2.

They also confirm that there will be two versions of Fiji, where the fastest one will only be sold by AMD.

http://www.hardware.fr/news/14241/computex-amd-fiji-ses-4-go-hbm-photo.html

 

[Vesku]

Hardware.fr confirms 4GB limit on Fiji, clamshell not possible width HBM1 but will be with HBM2.

They also confirm that there will be two versions of Fiji, where the fastest one will only be sold by AMD.

http://www.hardware.fr/news/14241/computex-amd-fiji-ses-4-go-hbm-photo.html

They get their dimensions from a partial render...

 

[RussianSensation]

AMD has artificially crippled FP64 on their consumer cards since hawaii so I don't see how this would matter at all for a consumer.

But will the FirePro Fiji XT be compute crippled? If not, AMD will have designed a 2-in-1 chip that's fast in both areas. GM200 can't claim that feat. That would mean AMD's engineers have outperformed NV's when taking things in the overall context of what the actual ASIC is capable of.

If you want a consumer card that's good at FP64 you buy tahiti at the lower end or Titan/Titan black at the higher end. Fiji is unlikely to change this.

That's not the point. Last gen FP64 was used to justify the huge premiums for the OG Titan. GM200 threw FP64 over the bridge and now 12GB is used to justify it, like taking a GTX580 3GB and nearly doubling its price.

Pretty much no one has said this since the synapse leaks showing 500mm^2+ nearly a year ago

You haven't followed the forums closely enough then. People did say AMD will not make a large die GPU this gen and that AMD's best card this series would just be either a rebranded Hawaii XT with HBM OR a dual-Tonga XT 2048 shader x 2 card. So yes, absolutely, it was being pushed for weeks how AMD's flagship card will never be a 500mm2+ 28nm chip.

Power consumption and efficiency remains to be seen. However it's going to be almost impossible to claim that the GCN in Fiji has better or worse perf/watt than any other AMD GCN part because HBM is going to completely ruin any comparisons between the architectural revisions.

What? I don't think you thought your response through. We compare perf/watt in 2 scenarios: videocard vs. videocard basis OR on a total system basis. No one here compares perf/watt of an ASIC itself without the PCB and its memory sub-system. This is something an hardware NV/AMD engineer or a fabrication engineer might do. I am scratching my head how you suddenly made up some new way of using perf/watt to discredit the improvements Fiji XT videocard will bring over R9 290X videocard on the same 28nm node.

 

[RussianSensation]

AMD seems to think it won't use more power than R290X, if the full German quote is taken and translated, it continues, saying that there's a very good chance it will use LESS power than R290X. To put that in context, non-reference R290X uses ~230-250W gaming load. Titan X uses ~235W gaming load.

No worries, the focus will shift to synthetic worthless FurMark numbers since it will be a guaranteed way to have a dual 8-pin videocard draw more power with a power virus than any 6-pin and 8-pin card.

While in the real world....

AMD should make an even better GPU call it the FURRY MAX PRO XT. They haven't use XT since Ati day though. Since it call furry, release it with a better TressFX technology as well.

They probably will in the forum of a dual Fiji Pro/Fiji XT R9 295X2 successor. It's a perfect fit for DX12 gaming since DX12 will allow VRAM stacking.

 

[3DVagabond]

You can see the core is raised up against the interposer at the top. I think those lighter/brown areas are the HBM

I think the black areas are the HBM.

 

[at80eighty]

"crushing" is such a subjective word, you'll find disagreement even among die hards.

also lol @ the laundry list, so people can save face

 

[antihelten]

No worries, the focus will shift to synthetic worthless FurMark numbers since it will be a guaranteed way to have a dual 8-pin videocard draw more power with a power virus than any 6-pin and 8-pin card.

While in the real world....

Interestingly enough if you take those numbers and the corresponding performance numbers, the efficiency advantage of the 980 over the 290X is only 53% and not the 91% reported by TPU.

Of course we're talking system power versus card power here, but the system would have to account for roughly 200 Watts to reach TPUs numbers, in which case you only have 75-90 Watts left for the 980 and 110-130 Watts for the 290X.

 

[RussianSensation]

Interestingly enough if you take those numbers and the corresponding performance numbers, the efficiency advantage of a system equipped with the 980 over a system equipped with the 290X is only 53% and not the 91% reported by TPU.

Of course we're talking system power versus card power here, but the system would have to account for roughly 200 Watts to reach TPUs numbers, in which case you only have 75-90 Watts left for the 980 and 110-130 Watts for the 290X.

Fixed.

Of course, but under no circumstances does NV marketing want any gamer to compare overall system efficiency in games in terms of perf/watt. That's why they create artificially made up TDPs of 145W/165W for GM204 and why sites like TPU are exclusively focusing in videocard perf/watt, ironic considering no videogame in the world can be played on just the videocard without a motherboard, VRAM, SSD/HDD and CPU present.

NV wants to spread marketing lies how a 970 uses 145W of power because that sounds a heck of a lot better against an R9 290X card with 300W TDP than seeing 40-50W of power usage differences in an i5/i7 gaming rig! Have you not noticed how perf/watt proponents keep using theoretical TDP measurements and refuse to compare overall system efficiency for gaming, like ever?!

 

[Cloudfire777]

Hardware.fr confirms 4GB limit on Fiji, clamshell not possible width HBM1 but will be with HBM2.

http://www.hardware.fr/news/14241/computex-amd-fiji-ses-4-go-hbm-photo.html

Thats what Ive been saying all along. AMD didnt defend 4GB just for fun

 

[caswow]

So guys the tone is set. 4gb is not good enough anymore.

 

[flopper]

So guys the tone is set. 4gb is not good enough anymore.

of course it is more than good enough.

 

[3DVagabond]

So guys the tone is set. 4gb is not good enough anymore.

We have to see the performance. We aren't comparing apples/apples. We're talking HBM to GDDR5. That's like saying you'd rather have 6GB of DDR3 instead of 4GB of GDDR5. Most would choose the latter.

 

[caswow]

We have to see the performance. We aren't comparing apples/apples. We're talking HBM to GDDR5. That's like saying you'd rather have 6GB of DDR3 instead of 4GB of GDDR5. Most would choose the latter.

yea but people jump on conclusions because its amd. if it was nvidia it woulnt count because its nvidia.

 

[DDH]

Tweaktown have spotted a powercolor 390x at computex with hybrid cooler. Looks interesting

http://www.tweaktown.com/news/45669/hands-one-powercolors-next-gen-radeon-r9-390x-video-cards/index.html