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

Esrever said:
I don't know how people started to think $650 is a price that is too low to make a profit.
Click to expand...
What is the yield of the 4096 core Fiji?
That chip is massive so with poor yields it will cost a bundle for the ones with no flaws.

Also, HMB yields are what?

What are the costs of the interposer?

Water cooling costs?

Adding those up could have the Fury at $650 being barely profitable and with the low volumes not doing much to AMD's bottom line.
 
Lanek makes for a very good case. Never once a WMD build (all 900 hundred of them) exhibited the issues observed on the retail build.


We know companies don't play fair
 
A1xLLcqAgt0qc2RyMz0y said:
What is the yield of the 4096 core Fiji?
That chip is massive so with poor yields it will cost a bundle for the ones with no flaws.

Also, HMB yields are what?

What are the costs of the interposer?

Water cooling costs?

Adding those up could have the Fury at $650 being barely profitable and with the low volumes not doing much to AMD's bottom line.
Click to expand...

Yields of HBM dont depend of the HBM production as it was a simple chips, it differ at every level, as it is a compilation of chips ... but i will ensure you that yields is not a problem.

- Yields on 28nm are really good, why ? well this make year TSMC do them now Yields and cost on 28nm are incredible, you nearly do money by product on their fabs right now ( sorry was a bit of the joke, ofc, it cost you money, but well, you have understand me )

- Interposer dont cost so much, it use recycled 65nm fabs.. The R&D have cost a lot, but the production cost are really, really low. They are really easy to do.

- HBM are using 22 stacked die, but in fact at the nm process, it cost nothing ( Samsung was using 65nm stacked die for showcase their 3D nand )


- WC cost is a good question, the system dont seem to be normal AIO, but part cost nothing... In a waterblock, what cost the most is the copper used for the coldplate . ..Technology behind H20 are purely industrial, no patent right, it is really simple, a pump is really simple,, a tube is a tube, a radiator is just made of copper tubes and metal fins .. it cost nothing ( i said that, but my complete watercooling system have cost me 1000$ but it is because brand like EK cant push the cost in big volumes )
 
Last edited:
A1xLLcqAgt0qc2RyMz0y said:
What is the yield of the 4096 core Fiji?
That chip is massive so with poor yields it will cost a bundle for the ones with no flaws.

Also, HMB yields are what?

What are the costs of the interposer?

Water cooling costs?

Adding those up could have the Fury at $650 being barely profitable and with the low volumes not doing much to AMD's bottom line.
Click to expand...

Fiji is a hair smaller than GM200 and that seems to be doing just fine on 28nm. The limiting factor would then be the interposer and HBM.
 
lanek said:
- Yields on 28nm are really good, why ? well this make year TSMC do them now Yields and cost on 28nm are incredible, you nearly do money by product on their fabs right now ( sorry was a bit of the joke, ofc, it cost you money, but well, you have understand me )

- Interposer dont cost so much, it use recycled 65nm fabs..

- HBM are using 22 stacked die, but in fact at the nm process, it cost nothing ( Samsung was using 65nm stacked die for showcase their 3D nand )


- WC cost is a good question, the system dont seem to be normal AIO, but part cost nothing... In a waterblock, what cost the most is the copper used for the coldplate . ..Technology behind H20 are purely industrial, no patent right, it is really simple, a pump is really simple,, a tube is a tube, a radiator is just made of copper tubes and metal fins .. it cost nothing ( i said that, but my complete watercooling system have cost me 1000$ )
Click to expand...

costs have a lot to do with supply and demand too, now HBM supply if low costs will be higher. The water cooling solutions supply will definitely be lower then the usual counterparts since these solutions aren't going to be as mass manufactured as the fan solutions.
 
11bp2c4.jpg


http://videocardz.com/56609/amd-radeon-r9-fury-x-r9-fury-x2-and-r9-nano-detailed-some-more
 
Razor1 said:
costs have a lot to do with supply and demand too, now HBM supply if low costs will be higher. The water cooling solutions supply will definitely be lower then the usual counterparts since these solutions aren't going to be as mass manufactured as the fan solutions.
Click to expand...

On this you are completely right, i mean the volume is not there for pull down the price ... this said, the cost of HBM was really on the developpement of it, a little bit like the Samsung 3D Nand, the production of it counter the cost of developpement, and you limit the risk .. Many parts of it is using "old " recycled fab ( the interposer in 65nm ), as an example, and in overall you obtain a real reduction of the cost of production... what can be risky when you dont have volume, the thing was to reduct the price of production, before the volume are here.. so on the end you obtain an new, innovative product, but who cost way less to product, even if the volumes are not there.. ( We can thanks Samsung for this lesson ).. And when the volume is here.. ( say hello to heaven ).

Peoples assume that HBM cost a lot, in fact HBM is made for reduct the cost when offer, innovative and better performance..
 
Last edited:
AlNets said:
How do they have mem clk listed for Nano and yet the effective clk is TBC?
Click to expand...

Because HBM memory clock is 500mhz whatever is the gpu ..and they just dont know other specification for it .. HBM 1.0 operate at 500mhz...

Oups sorry, i have misunderstand you .. effective clock will not change .
 
lanek said:
- HBM are using 22 stacked die, but in fact at the nm process, it cost nothing ( Samsung was using 65nm stacked die for showcase their 3D nand )
Click to expand...
Samsung's 3D NAND is not a stacked die. The storage rows are literally fabricated as vertical structures. This is mostly acceptable here because the actual activity level of NAND storage can accept exceedingly poor electrical performance relative to what having actual layers of silicon crystal would provide for logic or DRAM.
The pricing picture for that has taken some time to be fully acceptable, and it does not require a very involved mechanical grinding, aligning, and bonding process every single layer.
Given the effective height for some of these technologies, Samsung would notice the cost of having to thin and mount 32 silicon dies per NAND chip, and AMD would notice it as well with 128 storage dies working in perfect order.

UniversalTruth said:
http://videocardz.com/56609/amd-radeon-r9-fury-x-r9-fury-x2-and-r9-nano-detailed-some-more
Click to expand...
It looks like they are just extrapolating values from various sound bites, and I do not think some of the assumptions will hold.
The 275W figure was not given as a TDP, for example.
 
3dilettante said:
Samsung's 3D NAND is not a stacked die. The storage rows are literally fabricated as vertical structures. This is mostly acceptable here because the actual activity level of NAND storage can accept exceedingly poor electrical performance relative to what having actual layers of silicon crystal would provide for logic or DRAM.
The pricing picture for that has taken some time to be fully acceptable, and it does not require a very involved mechanical grinding, aligning, and bonding process every single layer.
Given the effective height for some of these technologies, Samsung would notice the cost of having to thin and mount 32 silicon dies per NAND chip, and AMD would notice it as well with 128 storage dies working in perfect order.


It looks like they are just extrapolating values from various sound bites, and I do not think some of the assumptions will hold.
The 275W figure was not given as a TDP, for example.
Click to expand...


Youb are right, technology differ, the words was for economic side, more than the technical ones .
 
silent_guy said:
Whenever AMD is about to release an top-end GPU, benchies show up that are very good. And then it's followed by disappointment. Maybe things are different this time, we'll know soon.

But if true, then I'm at a loss about the pricing strategy: AMD has just stated that they don't want to be in the position anymore of the guy you go to when you want a lower price.
They'd have better performance, they have a GPU that looks amazing, and they have the unique story of HBM.

The crowd that buys Titan X would eat up a FuryX at $850. With a good air cooler, they could still price the Fury at $650 and it would sell amazing.

I don't get it.
Click to expand...

There is also alternative approach to your logic.

The crowd who buys Titan X already bought Titan X. The crowd waiting for AMD is mostly waiting for better offer at similar performance. The crowd sitting on GTX 6xx 7xx and R9 28x and R9 290x would never ever have paid $850 for a card offering around 50% more performance than R9 290X currently costing $300 - $350.

Your pricing structure would make AMD look bad to their loyal customers while nVidia fans would still say 'oh, nice card but it's still AMD and (put your fav. word like drivers or noise or temps.) will suxx' and then ignore it. Then there is also the issue of 4GB RAM. It's a lot less than Titan X and still less than 980Ti so for Joe Average paying more for less is not a good deal!

I personally would not buy Fury X for $850, not with less than 2x performance of 290X. I'd rather go CF and suffer all it's pitfalls or wait for 16nm FinFET next year.
AMD did well pricing this line up and I think they considered all pros and cons more thoroughly than me or you ever will.

PS. Since when people want corporations to make money at their expense? What world we are living in!?!
 
Last edited:
lanek said:
Youb are right, technology differ, the words was for economic side, more than the technical ones .
Click to expand...
The economic side reflects the difference as well.
3D NAND increases the number of steps and complexity of the fabrication process, and the business case for that has only become strong enough rather recently for very cost-sensitive NAND.

Stacked dies involves multiple silicon dies, which means substrate costs are multiplied per layer, which is a cost adder. The actual mechanical steps and risks involved with getting dies to align and bond with tolerances measured in microns are a significant change, and happen per layer. HBM1 stopped at four layers, to Fury's disadvantage.
The TSVs necessary for this and the interposer itself are a cost adder. The interposer's TSVs are disliked enough that Intel and various others are working to find ways of avoiding them. It may not seem like much for a boutique video card purchase, but the manufacturers that have to live with it don't like it.
3D NAND is too different to serve as an example of why HBM doesn't face significant cost concerns versus other DRAM solutions.
 
Pressure said:
Fiji is a hair smaller than GM200 and that seems to be doing just fine on 28nm. The limiting factor would then be the interposer and HBM.
Click to expand...

Nvidia has made GPU's close to 600 mm2 for some time now and has got a handle on yields now but in the early days yields stunk.

Isn't this AMD's first attempt in making giant GPUs? First attempts seem to have low yields at the start then improve over time with tweaks.
 
Razor1 said:
its possible that video shows similar issues with PCcars as it does with Rise. The nV cards aren't as bottlenecked by the CPU.
Click to expand...

AFAIK, the issue with Project Cars is that it's using PhysX for car physics simulation (there's no other alternative than to use PhysX).
If you're using a nVidia card, it'll do PhysX through the graphics card. If you own an AMD/Intel card, it'll have to use the PhysX CPU path.

And the biggest problem here is the PhysX CPU path, which is purposely botched. IIRC, it's locked onto a single-thread and uses ~30 year-old x87 instructions for floating point operations instead of the much faster SSEx (yeah nVidia did a really good job at crapping over other IHVs again optimizing for their own hardware).

Had the game used another physics engine (e.g. Havok), none of this would've happened.
Which is probably why we have never seen PhysX being used for anything other than visual effects that you can switch on/off or some periodic and light physics simulations.
It's too bad that Slightly Mad Studios couldn't see this before the shit hit the fan.
Like I said, Gameworks deals are like peeing in your pants.

I've seen reports of a DX12 build of Project Cars running about twice as fast for AMD GPU users. Basically, freeing up CPU resources with the new API leaves more time for the CPU to run all those super-inefficient physics simulations.



To summarize:
- Terrible performance on Project Cars in AMD/Intel GPUs has little to do with AMD drivers. Don't use PhysX in simulations that you can't disable. Avoid Gameworks like the plague.
 
I was reading Anandtech news about dual Fiji card and from their picture you can read the ES chip markings. It was made on 47th week of 2014 (2nd half of November).
 
ToTTenTranz said:
AFAIK, the issue with Project Cars is that it's using PhysX for car physics simulation (there's no other alternative than to use PhysX).
If you're using a nVidia card, it'll do PhysX through the graphics card. If you own an AMD/Intel card, it'll have to use the PhysX CPU path.

And the biggest problem here is the PhysX CPU path, which is purposely botched. IIRC, it's locked onto a single-thread and uses ~30 year-old x87 instructions for floating point operations instead of the much faster SSEx (yeah nVidia did a really good job at crapping over other IHVs again optimizing for their own hardware).

Had the game used another physics engine (e.g. Havok), none of this would've happened.
Which is probably why we have never seen PhysX being used for anything other than visual effects that you can switch on/off or some periodic and light physics simulations.
It's too bad that Slightly Mad Studios couldn't see this before the shit hit the fan.
Like I said, Gameworks deals are like peeing in your pants.

I've seen reports of a DX12 build of Project Cars running about twice as fast for AMD GPU users. Basically, freeing up CPU resources with the new API leaves more time for the CPU to run all those super-inefficient physics simulations.



To summarize:
- Terrible performance on Project Cars in AMD/Intel GPUs has little to do with AMD drivers. Don't use PhysX in simulations that you can't disable. Avoid Gameworks like the plague.
Click to expand...

I believe lanek is talking about PC Cars not Project Cars. Unless he made a mistake of course.
 
You must log in or register to reply here.

Similar threads

U
AMD: Southern Islands (7*** series) Speculation/ Rumour Thread
236 237 238
Replies
5K
Views
1M
BrynS
B
Back
Top