• Competitor rules

    Please remember that any mention of competitors, hinting at competitors or offering to provide details of competitors will result in an account suspension. The full rules can be found under the 'Terms and Rules' link in the bottom right corner of your screen. Just don't mention competitors in any way, shape or form and you'll be OK.

AMD Polaris architecture – GCN 4.0

Mauller said:
Samsung has been producing 20nm class 4GB HBM2 modules since the begging of this year. Creating designs for 14 and 16nm straight away is just future proofing. HBM2 is already in production.

The later release of GP100 is due to dire yield for such a big part this early on, considering they also need a fully working 1200mm interposed. And even with Q1 17, that is still only to select HPC partners and Universitys.

You can even check the press releases on Samsung's own website.
Click to expand...

The problem is that if they were mass producing it for so long,then we would have seen even Vega being released together with Polaris,or even Polaris potentially having it. At the moment that would mean HBM2 production for over 9 months with no mass production customer as the GP100 is meant to not be mass availability for 9 months at least.

It is only strange they are shifting so quickly from 20NM production to BOTH 14NM and 16NM production. If anything that indicates 20NM HBM2 is having major issues I suspect,and the only reason Samsung is saying anything is to "show" they have production and one over SK Hynix.

Too many people are believing all the marketing bumpf.

OFC,AMD might throw a curb ball at us and have another higher tier graphics card using HBM2 in the meantime before Vega,but I do doubt it.
 
Last edited:
CAT-THE-FIFTH said:
The problem is that if they were mass producing it for so long,then we would have seen even Vega being released together with Polaris,or even Polaris potentially having it. At the moment that would mean HBM2 production for over 9 months with no mass production customer as the GP100 is meant to not be mass availability for 9 months at least.

It is only strange they are shifting so quickly from 20NM production to BOTH 14NM and 16NM production. If anything that indicates 20NM HBM2 is having major issues I suspect,and the only reason Samsung is saying anything is to "show" they have production and one over SK Hynix.

Too many people are believing all the marketing bumpf.

OFC,AMD might throw a curb ball at us and have another higher tire GPU using HBM2 in the meantime before Vega,but I do doubt it.
Click to expand...

I think you're making a lot of assumptions here. It's just as likely that we've not seen HBM2 cards because the new GPUs are not ready, or that the expense of new HBM2 memory isn't viable on a low/mid-range part (Polaris) that has to address a specific market segment/price.

The HBM2 roadmap referenced above gives more timescales about HBM2 availability than we've ever had from Nvidia/AMD about their GPU parts that will utilize HBM2. Yet you're discounting the above info in favour of the absence of info from Nvidia/AMD.
 
Mauller said:
Samsung has been producing 20nm class 4GB HBM2 modules since the begging of this year. Creating designs for 14 and 16nm straight away is just future proofing. HBM2 is already in production.

The later release of GP100 is due to dire yield for such a big part this early on, considering they also need a fully working 1200mm interposed. And even with Q1 17, that is still only to select HPC partners and Universitys.

You can even check the press releases on Samsung's own website.
Click to expand...

That information has precisely nothing to do with the memory chip production at all, only the phy layer. The key should be that it's esilicon slides and not Hynix or Samsung.

You don't make a different memory chip for every possible interposer/logic chip combination you just make standardised chips then you make specific PHY, physical layer interface chips. Either esilicon is making such chips or is providing the IP by working with various customers and processes to create a standardised IP everyone can buy to make sure everything works together.

Till that stuff is basically produced and tested you don't have the ability to mass produce chips that can work together. So it's a good indication of why you'll have mass produced products working with them as long as you know who they are working with. I don't know the industry well enough, maybe they are the go to company everyone works with, maybe they have several competitors. What it would seem from that presentation is that esilicon is fairly heavily involved in the HBM supply chain at least in terms of working with hynix and Armkor to package anything made from glofo/samsung and likely TSMC. It's likely AMD used esilicon for 28nm HBM1 along with Armkor for packaging and TSMC for the production of the chips.


As for interposers, 1200mm is nothing. Interposers won't be made on leading edge process nodes, there is no need as the circuitry isn't complex or dense so the cost for doing so is not worth it. Fiji's was made on 65nm as far as I know. Your average chip is made on the smallest node available, takes 4-8 weeks to make and has multiple layers, base metal layers which is mostly just traces/connections while higher layers are where you have transistors and complex structures. Interposers are just the base layers which are extremely easy and quick to make. If a 65nm wafer was low yield and cost $4-5k when new, today it's insanely high yield, probably 1-2k for a normal wafer and I would guess 1/3 of that cost at most for an interposer wafer. The logic is simple as well, HBM has redundant connections, the gpu will as well, to reroute data around any failed bumps or broken traces, with redundancy built in yields of even 1200mm interposers will be easily over 90% on a stupid cheap process with zero capacity constraints.

Interposers are a non issue and the phy is a step needed between memory production and actual ability to connect memory to chips designed to use them.
 
Steampunk said:
I think you're making a lot of assumptions here. It's just as likely that we've not seen HBM2 cards because the new GPUs are not ready, or that the expense of new HBM2 memory isn't viable on a low/mid-range part (Polaris) that has to address a specific market segment/price.

The HBM2 roadmap referenced above gives more timescales about HBM2 availability than we've ever had from Nvidia/AMD about their GPU parts that will utilize HBM2. Yet you're discounting the above info in favour of the absence of info from Nvidia/AMD.
Click to expand...

I am not really making that many - who is using HBM2?? AMD and Nvidia. Yet supposedly Samsung has been producing it since January. AMD Vega is meant to be an almost end of year release. The GP100 is not even going to enter mass availability until early 2017.

So that is Samsung producing HBM2 for 9 months plus with only a limited production GP100 taking some of it in 2H 2016.

Then within a year they are scrambling to get production on BOTH 14NM and 16NM ready and then even trying to see if they can use 28NM to produce the chips.

That looks like 20NM HBM2 is having problems to me. If it sounds like a Duck and walks like a Duck,it is a Duck.

I expect Vega and the mass production GP100 based cards to use 14NM and 16NM HBM2,or simply a case of stockpiling enough 20NM HBM2 so they can have enough for AMD and Nvidia.

On top of this,there are rumours Vega was delayed. AMD going by prior launches tends to not really go over 350MM2 on a new node and Nvidia is launching a 300MM2 GP104 which is similar in size. Yet the AMD part is launching AFTER the Nvidia part.

This has not happened for a very long time - AMD got the HD5000 and HD7000 out months before Nvidia even got their smaller parts released.

Nvidia has recently stopped production on both the GM200 and GM204 meaning during summer they intend to launch the replacement GP104.

However,it will be using GDDR5 and I suspect AMD has delayed Vega more due to HBM2 issues than the GPU itself.

I can't see Polaris 10 being faster than a Fury X too - unless OFC AMD have been doing some classic misinformation and have an HBM2 equipped Polaris 10 variant(or indeed a Polaris 9) waiting in the wings.
 
Last edited:
Steampunk said:
I think you're making a lot of assumptions here. It's just as likely that we've not seen HBM2 cards because the new GPUs are not ready, or that the expense of new HBM2 memory isn't viable on a low/mid-range part (Polaris) that has to address a specific market segment/price.

The HBM2 roadmap referenced above gives more timescales about HBM2 availability than we've ever had from Nvidia/AMD about their GPU parts that will utilize HBM2. Yet you're discounting the above info in favour of the absence of info from Nvidia/AMD.
Click to expand...

http://hexus.net/tech/news/graphics/79513-samsung-starts-mass-produce-8gb-gddr5-dram/

http://www.anandtech.com/show/10179/nvidia-announces-24gb-quadro-m6000

http://www.anandtech.com/show/10253/amd-announces-32gb-firepro-w9100

Actually people are reading too much into the dates given by Samsung.

Samsung started producing 8Gb chips of gddr5 last January. AMD launched the S9170 with 32GB of ram around last July but there has been next to zero availability till a month or so ago and AMD have also now launched the higher volume W9100 with 32GB in the past few days.

Drop in replacement gddr5 took 15 months from production starts to actual realistic availability(in what is still a lower volume market), but people think HBM2 production starting in Jan means it will be widely available within 6 months. These are facts, not assumptions, the assumptions being made about production dates through to large scale availability are the ones that are completely incorrect here.
 
CAT-THE-FIFTH said:
I can't see Polaris 10 being faster than a Fury X too - unless OFC AMD have been doing some classic misinformation and have an HBM2 equipped Polaris 10 variant(or indeed a Polaris 9) waiting in the wings.
Click to expand...

Its a strange one, as after the Hitman demo was shown, we had people saying it'll be about 15/20% faster than the Fury X, and now we have, it'll not be faster than the Fury X, as its expected to be around the 390s performance, so im just going to wait and see. :p
 
LoadsaMoney said:
Its a strange one, as after the Hitman demo was shown, we had people saying it'll be about 15/20% faster than the Fury X, and now we have, it'll not be faster than the Fury X, as its expected to be around the 390s performance, so im just going to wait and see. :p
Click to expand...

Well,I always said R9 390 to Fury level performance,so it wasn't me LOL! :p
 
Steampunk said:
I think you're making a lot of assumptions here.
Click to expand...


To be honest, all of us are. some of us are more knowledgeable about certain areas than others, but all any of us can do is look at the actual facts that we have and interpret them to give a reasonable estimation of what will happen.

What this info about eSilicon actually means, we don't really know as we have no idea if AMD or NVidia are even customers of theirs, who else might need HBM2 for something, I have no idea, but it seems that quite a lot of companies are making, or have plans to make the stuff just for it to be only used by NVidia and AMD on a couple of cards each and even then possibly much latter this year or even early next. (and yes that is more speculation and assumption)

Bottom line, none of us know much at all for certain, but if we didn't like the speculation and debate over it all we wouldn't post here in the first place.
 
I was going to wait till Polaris released their new GPU but couldn't resist and brought myself a R 390x Strix, my His 7950 will be missed lasted me 4 years such a beaut of a card.
 
LoadsaMoney said:
Its a strange one, as after the Hitman demo was shown, we had people saying it'll be about 15/20% faster than the Fury X, and now we have, it'll not be faster than the Fury X, as its expected to be around the 390s performance, so im just going to wait and see. :p
Click to expand...

It was said so because everyone thought Hitman was maxed out. Now we know it wasn't, so we can chill with all that 'ridicule' ;)
 
bru said:
To be honest, all of us are. some of us are more knowledgeable about certain areas than others, but all any of us can do is look at the actual facts that we have and interpret them to give a reasonable estimation of what will happen.

What this info about eSilicon actually means, we don't really know as we have no idea if AMD or NVidia are even customers of theirs, who else might need HBM2 for something, I have no idea, but it seems that quite a lot of companies are making, or have plans to make the stuff just for it to be only used by NVidia and AMD on a couple of cards each and even then possibly much latter this year or even early next. (and yes that is more speculation and assumption)

Bottom line, none of us know much at all for certain, but if we didn't like the speculation and debate over it all we wouldn't post here in the first place.
Click to expand...

Yeah, that's all we got is speculation....that's what makes me wonder why people get so caught up in it all and get quite confrontational about something they know very little about in all reality. No-one I know on here works for AMD or Nvidia (Opps Sorry AMDMatt, forgot about you for a minute) and even if they did would not be allowed to say anything anyway (AMDMatt hasnt said anything...thus proving my point). :p

Lets just wait for the cards to come out so we can have it all kick off for real in a few months eh! LOL Must agree though if we didnt like all the banter then we wouldnt be here, would we. :D
 
AMD’s Radeon 400 Series Polaris GPUs Land Major Apple Design Wins – Perf Per Watt & Perf Per Dollar Strides Pay Off


The Polaris 10 and 11 chips will go into new desktops and notebooks from Apple, which the company plans to bring to market later this year. And although these Apple design wins may not be significant volume contributors they are very profitable.

Latest Nintendo NX Rumours Suggest It Is Using Custom AMD Polaris 14nm GPU

We’ve known for some time now that AMD is working on whatever APU solution is used in Nintendo’s NX console, but it looks as if the Japanese giant could be taking advantage of the upcoming AMD die shrink. Polaris utilises the 14nm FinFET fabrication node and as such would likely be more powerful and much more energy efficient than the PlayStation 4.
 
Mtom said:
AMD’s Radeon 400 Series Polaris GPUs Land Major Apple Design Wins – Perf Per Watt & Perf Per Dollar Strides Pay Off


The Polaris 10 and 11 chips will go into new desktops and notebooks from Apple, which the company plans to bring to market later this year. And although these Apple design wins may not be significant volume contributors they are very profitable.

Latest Nintendo NX Rumours Suggest It Is Using Custom AMD Polaris 14nm GPU

We’ve known for some time now that AMD is working on whatever APU solution is used in Nintendo’s NX console, but it looks as if the Japanese giant could be taking advantage of the upcoming AMD die shrink. Polaris utilises the 14nm FinFET fabrication node and as such would likely be more powerful and much more energy efficient than the PlayStation 4.
Click to expand...

It would be a nice change if Nintendo had the most powerful console for a once.
 
Mtom said:
AMD’s Radeon 400 Series Polaris GPUs Land Major Apple Design Wins – Perf Per Watt & Perf Per Dollar Strides Pay Off


The Polaris 10 and 11 chips will go into new desktops and notebooks from Apple, which the company plans to bring to market later this year. And although these Apple design wins may not be significant volume contributors they are very profitable.

Latest Nintendo NX Rumours Suggest It Is Using Custom AMD Polaris 14nm GPU

We’ve known for some time now that AMD is working on whatever APU solution is used in Nintendo’s NX console, but it looks as if the Japanese giant could be taking advantage of the upcoming AMD die shrink. Polaris utilises the 14nm FinFET fabrication node and as such would likely be more powerful and much more energy efficient than the PlayStation 4.
Click to expand...

http://wccftech.com/nintendo-nx-gpu-power-rumors-debunked-by-insider/

it is, or it isn't

AMD are also already the only option in Apple's range, so I wouldn't pin my hopes on big sales numbers coming from that route either
 
Last edited:
You must log in or register to reply here.
Back
Top Bottom