And AMD will sell the defective ones as Fury and nano. So actually 3 SKUs. therefore one could argue they get even higher yield and hence cheaper than 980Ti going by the original argument.
In the huge rumor thread there was link to the company fabbing the interposer. ShintaiDK was all over that thread and he for sure saw that too as it was discussed. So yeah, now he is intentionally playing dumb. The cost of the interposer really is in the $3-4 range and will only get cheaper. AFAIK it's fabbed at 65 nm and you only need the cheapest few steps of that process. it's not at all comparable to fabbing a CPU or GPU.
On the downside though, it is the bill of materials that is the biggest question hanging over HBM. Since HBM introduces several new technologies there are any number of things that can go wrong, all of which can drive up the costs. Of particular concern is the yield on the HBM memory stacks, as the TSV technology is especially intricate and said to be difficult to master. The interposer on the other hand is simpler, but it still represents something that has never been done before, and AMD admits upfront that the manufacturing facilities being used to create the interposer are old 65nm lines originally used for full chip production. So while the interposer does not approach the cost of a full logic chip, there is still the matter of the existing manufacturing lines being sub-optimal for high-volume low-cost production. Meanwhile AMD does get to enjoy some cost savings as well the HBM PHYs are certainly much easier to implement than GDDR5 PHYs on Fiji itself, and the overall package is cheaper since it doesn't have GDDR5 memory running through it though it's unlikely that these savings outweigh the other costs of implementing HBM at this time.
Ultimately AMD Is not willing to discuss HBM costs or yields at this time. Practically speaking its not a consumer matter what matters to video card buyers is the $650 price tag on the R9 Fury X and from a trade secrets perspective AMD is loath to share too much about what they have learned since they are the first HBM customer and want to enjoy as much of that advantage as is possible. At this point I feel its a safe bet that the 4GB HBM implementation on Fiji is costing AMD more than the 4GB (or even 8GB) GDDR5 implementations on Hawaii cards, but beyond that its difficult to say much more on costs.
Would really suck if with current setup, the cost saving from using HBM over GDDR5 (and all the perks such as smaller PCB, less chips, less silicon/lines/etc) is offset by the cost of the interposer.
There always has to be a "but", doesn't there?
HBM currently cost more than GDDR5. I wouldnt be surprised if it was 2x. Interposer etc is just another part of its own.
I dont think HBM will be cheaper than GDDR5 before sometime in 2018 if I had to guess. It takes 1-2 years for DDR4 to get cheaper than DDR3 depending on how optimistic you see it. And I think its safe to say it wasnt 1 year.
Well I was looking specifically at this design and the overall (ie everything that ties to make Fiji). I'm sure HBM on itself cost more to make than GDDR5 (that I never actually doubted).
But I was under the impression that with the other cost savings coming with HBM that it would either balance out or not be a complete upset. If the overall cost is staggering, it is a woof.
EDIT: IE taking you're wording, if for example they had designed Fiji to use GDDR5 (ignoring the other limitations, such as power draw, etc etc) the total cost per chip would be half, that is a huge upset. I assumed it was, perhaps between 0-15% more factoring in all the other benefits.
There isnt really any other cost savings. PCB is pennies.
HBMs primary goal is speed and power consumption. I dont think anyone ever thought about it as cheaper as a goal.
Again, there are other factors (I guess thinking about it more now so, using the AIO CLC blew that out of the water haha) I was also thinking such things as less material for coolers, shrouds, less silicon lines, less voltage regulation, etc, etc. Small little things that just add up.
The difference between using HBM and GDDR5 isn't just "PCB" size/difference.
What a joke. AMD quite obviously competes with nVidia. Their last single gpu is within a few percent at 4k. And they still have the fastest single card (295x2). Fury X doesnt win generally, but that doesnt mean its not competitive. A price drop and it would be highly competitive. AMD's CPUs on the other hand are not competitive. In short, no, you're 100% wrong.
Everyone has a given right to be emotional about hardware and to act like the name of the desired hardware would better fit to them with such high expectations. 30% is like a doubling of the 295 performance, isn't it?https://www.youtube.com/watch?v=9l-7kFsqNjU
According to this reviewer, most games are faster on the 980ti . I feel stupid now waiting for this card that I was hoping to be at least 15 to 30% faster at the same price point or less than the 980ti. Not only does it not come with HDMI 2.0, it doesn't even come with 6 GB or more of HBM.
There are few AMD user left and this was the final nail in the coffin on how I feel about AMD. This gpu is going to alienate the very few AMD user left who were all hoping AMD would hit the ball out of the park with this one.
I have defect over to Nvidia now. No amount of right could change my mind now. AMD had a chance to make an epic gpu and they end up with another failure that also require a large cooler. Might as well make the GPU market a monopoly because AMD are either too cheap or they just can't complete. Their cpu is also a failure as well.
I feel very foolish for not listening to other people who said that the 980ti was better than the AMD Fury X and the wait wasn't worth it.
This is my rant, waited all of this time for an inferior product. Seeing how Eyefinity is important for me and is really the only reason I want to upgrade in the first place, I don't even know how I feel about building a PC now. This is utter disappointment. I can't voice my displeasure enough.
What is the power consumption then? In one review they added 200MHz to the Ti's GPU and 400MHz to memory. This increased system power by 46W. The Fiji was stable at a 63MHz OC, increasing power by 20W.But if you take the average 980 Ti overclock Boost level of 1400MHz+ and compare to the average Fury X OC of +100MHz (at best, some are doing as poorly as +50MHz max), the result is the Ti is +20% faster, in-game, using Anand's benchmark suite, versus the FX, OC v. OC.
[...]
Just a lone example, that Zotac Amp Extreme 980 Ti overclocked today in the Guru3D review to 1550MHz...just an incredible beast, obviously not the norm..and in those benchmarks it makes the FX look like a generation-old card.
What is the power consumption then? In one review they added 200MHz to the Ti's GPU and 400MHz to memory. This increased system power by 46W. The Fiji was stable at a 63MHz OC, increasing power by 20W.
980TI is a cutdown GM200 part.
Fury X is a full Fiji part.
The Fury that will sell for 550$ is equal to the 980TI in that area. And nVidia gets 100$ more per card there.
Feel free to prove it. Everyone says the interposer is (relatively) expensive. The 3-4$ number comes from Abwx doesnt it?
A 300mm interposer wafer cost between 500 and 650$ it seems. How many Fiji interposers can you fit on a 300mm wafer?
http://www.ocaholic.ch/uploads/extgallery/public-photo/medium/amd-fijigpuvc-2_674_1e071.jpg[/IMG[/QUOTE]
you can fit about 89 on a 300mm^2 wafer. Assuming the yields on it must be new 100%. simple maths gives you $7.3 per. Far from expensive. Shouldnt get costly overall.
I would guess the PCB might cost more, but the fury also has a PCB, even if its simpler.
you can fit about 89 on a 300mm^2 wafer. Assuming the yields on it must be new 100%. simple maths gives you $7.3 per. Far from expensive. Shouldnt get costly overall.
I would guess the PCB might cost more, but the fury also has a PCB, even if its simpler.
you can fit about 89 on a 300mm^2 wafer. Assuming the yields on it must be new 100%. simple maths gives you $7.3 per. Far from expensive. Shouldnt get costly overall.
I would guess the PCB might cost more, but the fury also has a PCB, even if its simpler.
HBM currently cost more than GDDR5. I wouldnt be surprised if it was 2x. Interposer etc is just another part of its own.
I dont think HBM will be cheaper than GDDR5 before sometime in 2018 if I had to guess. It takes 1-2 years for DDR4 to get cheaper than DDR3 depending on how optimistic you see it. And I think its safe to say it wasnt 1 year.
I think this is a poor analogy.
DDR4 is not really that substantial an improvement over DDR3; in most applications it will do very little, and it's currently supported only on expensive servers and HEDT products. It will likely become mainstream (i.e. reach price parity with DDR3) once Skylake comes out and it starts shipping in high-volume mainstream systems.
HBM/HBM2 is more technically difficult to implement, but the potential rewards for doing so are also greater. In mobile applications, performance/watt is extremely important, and HBM can improve that substantially. Thus, there is going to be a push to get HBM into smaller mobile-focused GPUs as soon as it is technically feasible.
My best guess is that HBM/HBM2 remains a premium product through 2016, but in early to mid 2017, it will become more mainstream. That's when the current plans have it being incorporated into APUs.
There isnt really any other cost savings. PCB is pennies.
HBMs primary goal is speed and power consumption. I dont think anyone ever thought about it as cheaper as a goal.
Seriously? I doubt Hynix wants this to remain a low-volume product for years. Mass production to get the price point more mainstream while maintaining profitability is definitely going to be part of their plans for the next couple of years.
As for the interposer, the current production (on old 65nm IC fab equipment) is basically a prototype. There are expected to be dedicated interposer production lines in the next couple of years, which will be cheaper since they will cut out the unnecessary part of the fabrication process.
89?
Try 48 if we assume its a completely square die. That leaves you at a best case scenario at just over 10$. And a worst case just over 13.5$.
Remember the wafer is round, and the bigger chips/interposers the bigger the waste area.
The interposer is cost wise one of the dark shadows of HBM. And we will see what relevance it holds in the future on products, if any.
HMC needs PCB I/O. Well, instead of comparing solutions in cases, where they obviously are a bad fit, we should look, what fits best.You say DDR4 is not important, while you talk about the performance/watt benefits for HBM. What is the main focus of mobile and servers?
Also if you payed any attention to HBM etc. All they talk about is speed/watt. Not density/watt. Take a wild guess on why. A 4GB 128GB/sec HMC module for example uses twice the power as a 4GB DDR4-2667.
I thought the times of production lines in those fabs are gone thanks to the modern transport systems and lots of automation. That should also be the key to provide flexible solutions to many different customers.As for the interposer, the current production (on old 65nm IC fab equipment) is basically a prototype. There are expected to be dedicated interposer production lines in the next couple of years, which will be cheaper since they will cut out the unnecessary part of the fabrication process.