Rumor: AMD gets priority advantage to Hynix HBM2

[csbin]

http://www.guru3d.com/news-story/rumor-amd-gets-priority-advantage-to-hynix-hbm2.html

This is an interesting story, apparently AMD secured the luxury of getting exclusive access to Hynix HBM2 graphics memory. The second iteration of High bandwith memory (stacked memory to be used on-die on the GPU) would be limited in its availability, AMD this way can secure exclusive access to it.

The rumor comes from wccftech though, so that is a bit icky as any news is news for that website, whether it's wrong or right doesn't matter. Based on a claim from 'sources within AMD' AMD is trying to get their hands on pretty much all HBM2 memory to block Nvidia. Nvidia will be using HBM2 with Pascal GPUs that should be released in 2016.

It also isn't sure if Hynix indeed is the company involved, however since Hynix is the only party making 3D stacked HBM graphics memory, that kinda speaks for itself. The current limitation for HBM graphics memory basically is the clock frequency, the number of stacks and thus the 4 GB limitation. Foir HBM2 that will change ,more stacks means more memory and the bandwidth in gen 2 would be higher as well. looking at some documents, the memory bandwith will be doubled up and HBM2 will allow for a total of 8GB of graphics memory.

 

[Sabrewings]

HBM has been an open standard since Ocotober of 2013. By next summer there will be other sources available as the other memory manufacturers get on board with their own lines.

 

[Flapdrol1337]

wccftech...

 

[shady28]

The problem AMD is having has to do with running Hawaii+HBM (Fiji) GPU. The fact that 384bit bus 980 Ti generally whips Fiji at 1440p and under shows that the Maxwell GPU itself is far superior. "Fiji" starts to win when Maxwell goes memory bandwidth limited. That's really easy to see when you look at the places where 980 Ti wins vs Fiji and vice versa. The same is true of the 256bit 980 vs 512bit 390X.

So the path AMD takes is to try to prevent Nvidia from getting a-hold of HBM? If AMD were really competent / capable you would think they would just make a better GPU to make better use of HBM, but no AMD will try to play supply chain games instead. I guess we will see how that works out for them.

 

[TemjinGold]

So nVidia is fine to do this with Gameworks but it's foul when AMD does this?

 

[shady28]

So nVidia is fine to do this with Gameworks but it's foul when AMD does this?

The proper analogy would be Gamworks vs Mantle. And the right way for both companies to do this if they don't like the API Microsoft provides would be to setup a a standards committee that they both sat on, and create standard extensions to MS DirectX that both can use.

TressFX doesn't count because AMD gets first dibs on what its design is. The fact that AMD is losing this battle overall simply points out the obvious - with proprietary APIs there will be a winner, and there will be a loser.

But that has nothing to do with trying to monopolize hardware components in the supply chain.

 

[Kenmitch]

HBM has been an open standard since Ocotober of 2013. By next summer there will be other sources available as the other memory manufacturers get on board with their own lines.

Yep....But doesn't mean jack if the best scenario uses/implementation of it are patented.

Theoretically the basic implementation and use could be covered by frand agreements. Doesn't mean every conceivable way of use is also included.

 

[wand3r3r]

Personally I doubt there will be a long period of exclusivity to AMD.

 

[The Alias]

The proper analogy would be Gamworks vs Mantle. And the right way for both companies to do this if they don't like the API Microsoft provides would be to setup a a standards committee that they both sat on, and create standard extensions to MS DirectX that both can use.

TressFX doesn't count because AMD gets first dibs on what its design is. The fact that AMD is losing this battle overall simply points out the obvious - with proprietary APIs there will be a winner, and there will be a loser.

But that has nothing to do with trying to monopolize hardware components in the supply chain.

However mantle ended up being the base for both dx12 and vulkan which ends up helping consumers

 

[monstercameron]

The problem AMD is having has to do with running Hawaii+HBM (Fiji) GPU. The fact that 384bit bus 980 Ti generally whips Fiji at 1440p and under shows that the Maxwell GPU itself is far superior. "Fiji" starts to win when Maxwell goes memory bandwidth limited. That's really easy to see when you look at the places where 980 Ti wins vs Fiji and vice versa. The same is true of the 256bit 980 vs 512bit 390X.

So the path AMD takes is to try to prevent Nvidia from getting a-hold of HBM? If AMD were really competent / capable you would think they would just make a better GPU to make better use of HBM, but no AMD will try to play supply chain games instead. I guess we will see how that works out for them.

did you forget your /s tag to denote sarcasm, if not this post is inaccurate to say the least.

 

[DeathReborn]

I thought HBM2 went up to 32GB? This story is so full of holes, inconsistencies and lies it must be April Fools Day all over again.

If this is true (which I highly doubt) wouldn't this kind of dirty deal make AMD a very naughty boy? (ref: Monty Python - Life of Brian)

 

[Sabrewings]

I thought HBM2 went up to 32GB?

I think it's up to 32Gb stacks, which would be 16GB. I do remember seeing 16GB listed on an Nvidia slide back in March.

 

[ShintaiDK]

With 4 stacks and 8hi stacks you get 32GB. Asuming Hynix doesnt mess up again.

But its safe to say nextgen GPUs will be 8-16GB and Pro cards 16-32GB.

 

[Azix]

I thought HBM2 went up to 32GB? This story is so full of holes, inconsistencies and lies it must be April Fools Day all over again.

If this is true (which I highly doubt) wouldn't this kind of dirty deal make AMD a very naughty boy? (ref: Monty Python - Life of Brian)

Considering nvidias stupid statements about how they did the work and why should they be expected to share it, they should be screwed on this in every way possible. By every company whose work they ever use.

 

[Sabrewings]

By every company whose work they ever use.

Got some examples of things they used? If any of them are a JEDEC standard, that's kind of (in a big way) the point.

 

[cbrunny]

Why would this surprise anyone?

 

[Azix]

Got some examples of things they used? If any of them are a JEDEC standard, that's kind of (in a big way) the point.

funny a company so against sharing work would have anything to do with jedec. even more hilarious said company uses vram technology developed by the competition for several generations.

they should get what they deserve. being shut out

They need to be put in timeout for a year or so to reconsider their stance against what makes the platform they participate in what it is. Getting a soft shutout from HBM is a good start

 

[chimaxi83]

Yawn. Wake me up when this matters to those of us who aren't diehard cheer leaders to either of these companies.

 

[TemjinGold]

The proper analogy would be Gamworks vs Mantle. And the right way for both companies to do this if they don't like the API Microsoft provides would be to setup a a standards committee that they both sat on, and create standard extensions to MS DirectX that both can use.

TressFX doesn't count because AMD gets first dibs on what its design is. The fact that AMD is losing this battle overall simply points out the obvious - with proprietary APIs there will be a winner, and there will be a loser.

But that has nothing to do with trying to monopolize hardware components in the supply chain.

I don't see why this is the proper analogy. With Gameworks, nVidia has exclusive access to black box code that gives them a massive advantage at launch. Now (if this story is true) AMD through exclusive access will (supposedly) have a massive advantage at launch. Both cases involve resources and both have the potential for a crippling advantage. The only difference is that one is software and one is hardware.

I'm not saying either action is proper. What I am saying is if nVidia is allowed to "play dirty," I don't see why AMD shouldn't be allowed the same. Nvidia gained their Gameworks advantage by investing resources into it. AMD is now gaining the same through the same means.

 

[Silverforce11]

TressFX doesn't count because AMD gets first dibs on what its design is.

AMD designed HBM. 2 years later, they invited Hynix to be their production partner.

JEDEC requires that members do not patent the technology and make it "open".

But! AMD owns patents regarding the memory controller for HBM. That's the real kicker, without a way to use HBM, the HBM chips themselves may be "JEDEC" open, it's useless.

http://www.uspto.gov/web/patents/patog/week11/OG/html/1412-3/US08984368-20150317.html

So hopefully NV was aware of it and have their own MC for HBM2 ready. I wonder how easy it is to implement a unique MC for HBM that will not infringe on AMD's patent.

 

[sandorski]

There is a difference between "Priority" and "Exclusive". Priority would mean they have first dibs on it, meaning that if AMD wasn't buying it all, the remaining could be sold to others. Exclusive would mean that only AMD gets access to whatever Hynix produces, even if they produced a large surplus.

Given that, if Hynix has poor supply, AMD could end up buying all of it under Priority, thus making somewhat equivalent to having Exclusive access. If Hynix has surplus(exceeds AMDs needs) supply, then Priority access is nothing like Exclusive access.

All that said, HBM might be difficult to Produce at this time with no solution in sight, yet anyway. If that's the case, Nvidia or others interested in using HBM might need to arrange another Manufacturer to produce HBM for them. It being an open Memory standard would allow them to do this. Other Manufacturers might have more difficulty than Hynix or it's possible that they will solve the issue on yields before Hynix.

As for the possibility that AMD might be using HBM in a way to lock out Nvidia from competition through nefarious means. Some defending Nvidia's similar actions have stated outright that AMD should do just that. So if that's what they are doing, I am 110% confident they will praise this action from AMD as the greatest act within PC gaming in years!

 

[Silverforce11]

Well NV had a free ride with GDDR3 and GDDR5... maybe AMD finally smartens up and ask them to buy a ticket for HBM2.

 

[Keysplayr]

There is a difference between "Priority" and "Exclusive". Priority would mean they have first dibs on it, meaning that if AMD wasn't buying it all, the remaining could be sold to others. Exclusive would mean that only AMD gets access to whatever Hynix produces, even if they produced a large surplus.

Given that, if Hynix has poor supply, AMD could end up buying all of it under Priority, thus making somewhat equivalent to having Exclusive access. If Hynix has surplus(exceeds AMDs needs) supply, then Priority access is nothing like Exclusive access.

All that said, HBM might be difficult to Produce at this time with no solution in sight, yet anyway. If that's the case, Nvidia or others interested in using HBM might need to arrange another Manufacturer to produce HBM for them. It being an open Memory standard would allow them to do this. Other Manufacturers might have more difficulty than Hynix or it's possible that they will solve the issue on yields before Hynix.

As for the possibility that AMD might be using HBM in a way to lock out Nvidia from competition through nefarious means. Some defending Nvidia's similar actions have stated outright that AMD should do just that. So if that's what they are doing, I am 110% confident they will praise this action from AMD as the greatest act within PC gaming in years!

Well, they have to do something. May as well be this. Imagine if this was true and Nvidia actually licensed HBM2 from AMD/Hynix?

 

[Silverforce11]

Well, they have to do something. May as well be this. Imagine if this was true and Nvidia actually licensed HBM2 from AMD/Hynix?

Unless they come up with their own unique MC for HBM (that won't infringe AMD's patents), they may have to do that.

Imagine that, each NV SKUs sold, AMD earns some $.

 

[Sabrewings]

I'm not entirely convinced that patent is for HBM. Note that it talks about parity data being stored. To my knowledge HBM doesn't do that and by avoiding that part Nvidia would not be violating this particular patent.

Also, it references chip stacks but nowhere does it say "HBM" or "high bandwidth memory." It's supposition that this applies to HBM and it keeps being touted as something it is not definitively.

Edit: By not doing ECC Nvidia would have nothing to worry about here. Does anyone even read these?

"6. The integrated circuit of claim 1, wherein: said reliability data comprises a plurality of error correcting codes (ECCs) including at least one ECC for each of said data elements."