[Micron]GDDR5X has Arrived

[PPB]

16Gbit ICs sound good to fight HBM's inherent advantages in the PCB size front.

Hopefully we can see 980ti+ performance and 8GB GDDR5 in reference ref960/ref970/Nano sized PCBs.

 

[dark zero]

So GDDR5X seems to be a competition of HBM... however it will get heavily outclassed by HBM2.

 

[ShintaiDK]

So GDDR5X seems to be a competition of HBM... however it will get heavily outclassed by HBM2.

GDDR5X solves the cost issue (and possible still density issue). As long as its there, HBM will never go anywhere broad. Hence HBM2 for flagship only.

 

[raghu78]

GDDR5X solves the cost issue (and possible still density issue). As long as its there, HBM will never go anywhere broad. Hence HBM2 for flagship only.

Keep dreaming. HBM2 will find its way into GPUs and APUs (both consumer and server). APUs will be even higher volume than discrete GPUs. So definitely HBM2 will find broad adoption. Its going to launch in 2017 in Vega 11 and Vega 10 and Zen APUs for server and consumer markets. I am sure by the time 2019 comes around and we see TSMC 7nm GPUs HBM2 will find its way down the product stack into higher volume. The big benefit is HBM2 will give power and form factor advantages over GDDR5X which are very useful in notebooks.

 

[ShintaiDK]

Keep dreaming. HBM2 will find its way into GPUs and APUs (both consumer and server). APUs will be even higher volume than discrete GPUs. So definitely HBM2 will find broad adoption. Its going to launch in 2017 in Vega 11 and Vega 10 and Zen APUs for server and consumer markets. I am sure by the time 2019 comes around and we see TSMC 7nm GPUs HBM2 will find its way down the product stack into higher volume. The big benefit is HBM2 will give power and form factor advantages over GDDR5X which are very useful in notebooks.

So now also HBM2 for consumer APUs too in 2017? I say you are a tad overly optimistic again.

As long as HBM cant compete in cost and density it will be limited. No matter how good it is.

 

[raghu78]

So now also HBM2 for consumer APUs too in 2017? I say you are a tad overly optimistic again.

As long as HBM cant compete in cost and density it will be limited. No matter how good it is.

AMD designed HBM and HBM2 as they knew thats the only way to address the bandwidth needs of their extremely powerful APUs. A 4C/8T Zen APU with a 1024 sp Vega GPU is going to badly bottlenecked by dual channel DDR4. Anyway I think AMD already mentioned last year that HBM will make its way into APUs. Combine that with Raja Koduri's statements where he said they want to introduce HBM2 when the pricing is good for mainstream products. I can see a 200-220 sq mm Zen APU being sold for USD 300 - 350. Similarly a Vega 11 GPU with HBM2 at 300-330 sq mm could also end up at USD 350 - 400. In fact the APU will actually have better margins. So I don't see it being a cost problem. The volume with APUs will also help with economies of scale and help bring HBM2 costs lower. The key here is Zen will make the CPU portion much more performance competitive with Intel SoCs thus making their APUs more attractive. Right now AMD APUs suffer from the stigma of Bulldozer legacy which makes them impossible to sell at higher ASPs.

 

[ShintaiDK]

Right...

Raja also said that Polaris series wouldn't get HBM due to cost.

 

[raghu78]

Right...

Raja also said that Polaris series wouldn't get HBM due to cost.

But he did not say Zen APUs won't get HBM2. I think the HBM2 production timelines are not aligned with Polaris launch schedules. Yield,cost and volume are also going to be more favourable in 2017 as Raja hinted with Vega 10/11 . Samsung started HBM2 production in Q1 2016. Hynix is going to start in Q2 2016 with availability in Q3 2016.

http://www.kitguru.net/components/g...mory-chips-opens-way-for-32gb-graphics-cards/

http://www.kitguru.net/components/g...mory-chips-opens-way-for-32gb-graphics-cards/

http://www.techpowerup.com/220707/sk-hynix-to-ship-4gb-hbm2-stacks-by-q3-2016.html

This gives them a year's production to work up yield and volume and bringing costs lower before we see Zen APUs with HBM2 in Q3 2017.

 

[ShintaiDK]

You think HBM2 is going to be cheaper than HBM1? And cheap enough to make it work. Including interposer and interposer size. Can the AM4 socket even contain it? Both size and power delivery with dual memory controllers. Not to mention the economics behind.

My guess is your hopes are completely fantasy driven.

It wasn't long ago someone predicted Polaris 10 would be HBM based too.

 

[jpiniero]

Yeah, I wouldn't even expect 2017 dGPUs to be HBM2 except at the super high end ($800+). It'll be GDDR5X. Maybe 2018.

 

[raghu78]

You think HBM2 is going to be cheaper than HBM1? And cheap enough to make it work. Including interposer and interposer size. Can the AM4 socket even contain it? Both size and power delivery with dual memory controllers. Not to mention the economics behind.

My guess is your hopes are completely fantasy driven.

It wasn't long ago someone predicted Polaris 10 would be HBM based too.

http://wccftech.com/amd-am4-%C2%B5opga-socket-1331/

The AM4 socket should not have any problems with power delivery given the 1331 pins. As for interposer size a 200-220 sq mm die with a single HBM2 stack is not going to be space constrained. My guess is based on Joe Macri's statements that HBM will make it to APUs.

http://www.anandtech.com/show/9266/amd-hbm-deep-dive/4

"Meanwhile shifting gears towards the long term, high-end GPUs are just the first of what AMD expects to be a wider rollout for HBM. Though AMD is not committing to any other products at this time, as production ramps up and costs come down, HBM is expected to become financially viable in a wider range, including lower-end GPUs, HPC products (e.g. FirePro S and AMD’s forthcoming HPC APU), high-end communications gear, and of course AMD’s mainstream consumer APUs. As lower-margin products consumer APUs will likely be among the farthest off, however in the long-run they may very well be the most interesting use case for HBM, as APUs are among the most bandwidth-starved graphics products out there"

Its a matter of timing (14nm Zen APUs) and cost/volume/yield reasons with HBM2. Zen addresses the weak single thread performance and power inefficiency problem of current AMD APUs based on Bulldozer legacy and allows them higher ASPs and room to afford costs with HBM2. Anyway I know you will not agree. Lets wait and see how the next 15-18 months play out.

 

[dark zero]

So now also HBM2 for consumer APUs too in 2017? I say you are a tad overly optimistic again.

As long as HBM cant compete in cost and density it will be limited. No matter how good it is.

Same with edRAM, sorry to shattering your dreams to see a nVIDIA powered card with edRAM as principal memory.

 

[Silverforce11]

But he did not say Zen APUs won't get HBM2.

It's because folks can't get rid of their pre-conceptions that an APU must be cheap and low-end only.

 

[ShintaiDK]

http://wccftech.com/amd-am4-%C2%B5opga-socket-1331/

The AM4 socket should not have any problems with power delivery given the 1331 pins. As for interposer size a 200-220 sq mm die with a single HBM2 stack is not going to be space constrained. My guess is based on Joe Macri's statements that HBM will make it to APUs.

http://www.anandtech.com/show/9266/amd-hbm-deep-dive/4

"Meanwhile shifting gears towards the long term, high-end GPUs are just the first of what AMD expects to be a wider rollout for HBM. Though AMD is not committing to any other products at this time, as production ramps up and costs come down, HBM is expected to become financially viable in a wider range, including lower-end GPUs, HPC products (e.g. FirePro S and AMD’s forthcoming HPC APU), high-end communications gear, and of course AMD’s mainstream consumer APUs. As lower-margin products consumer APUs will likely be among the farthest off, however in the long-run they may very well be the most interesting use case for HBM, as APUs are among the most bandwidth-starved graphics products out there"


Its a matter of timing (14nm Zen APUs) and cost/volume/yield reasons with HBM2. Zen addresses the weak single thread performance and power inefficiency problem of current AMD APUs based on Bulldozer legacy and allows them higher ASPs and room to afford costs with HBM2. Anyway I know you will not agree. Lets wait and see how the next 15-18 months play out.

Just because it got 1331 pins means nothing in relation to HBM. Unless you can show the existence of specific HBM purpose pins. These will have to exist from day 1 with Bristol Ridge launch. So we dont have to wait that long. And just to show how silly it is. My 6700K got 1151 pins, the same chip in a laptop got 1440 pins. Yet there is no secret or magic hidden.

So you think the timeline he refers to is going to be 2017. You may need quite a bit more patience.

Not to mention the news comes from "Fiji is an overclockers dream". And you base it on what he didn't say.

Same with edRAM, sorry to shattering your dreams to see a nVIDIA powered card with edRAM as principal memory.

I think you are the only person in the world with that idea.

 

[Adored]

Zen APU in 2017 will have HBM. There are at least two obvious reasons here that nobody seems to have realised, but when you do you'll realise why HBM was the only choice.

 

[jpiniero]

It's because folks can't get rid of their pre-conceptions that an APU must be cheap and low-end only.

OEMs won't buy it unless it's cheap. Seriously.

 

[raghu78]

It's because folks can't get rid of their pre-conceptions that an APU must be cheap and low-end only.

exactly. The problem with AMD APUs was the CPU performance was too weak against Intel architectures (Sandy/Ivy/Haswell/Broadwell/Skylake) due to the Bulldozer disaster. The APUs could not be sold at prices of USD 300-350 like Intel quad cores because Bulldozer was extremely power and area inefficient. Zen is designed to address that problem. A 4c/8T Zen APU with competitive perf and efficiency should help AMD APUs to sell at higher prices of USD 300-350. btw TSMC 7nm has a high performance node

http://www.eetimes.com/document.asp?piddl_msgid=356922&doc_id=1329217&page_number=2

"The 7nm platforms will have their own development kits, EDA tools and IP blocks. In addition, companie a s such as ARM and Imagination Technologies will provide IP reference subsystems and optimized processors to speed SoC integration and verification, said Hou.

The HPC variant will make chips that run 10-15% faster than those made in the mobile process, enabling clock trees optimized to hit 4 GHz data rates. The variant will include optimized SRAM compilers supporting L3 caches and taller standard cell libraries. Early versions of the tools and IP blocks will be available later this year."

AMD APUs built at TSMC 7nm or equivalent process nodes from Samsung/GF could face off against Intel 10nm chips in 2019. So we are looking at Polaris 10 GPU (or even better) in a APU in 2019. AMD has a good opportunity to compete with Intel on process nodes competitive to Intel which has never happened in the past.

 

[3DVagabond]

OEMs won't buy it unless it's cheap. Seriously.

This is the future. Discrete boards, discrete ram, discrete SSD, are all going to go away.

 

[raghu78]

OEMs won't buy it unless it's cheap. Seriously.

I disagree. OEMs will buy it if AMD provides competitive CPU perf and GPU performance comparable to discrete GPUs of similar size. In fact I would say AMD APUs will be a better option than a Intel SoC with nVidia GPU in power constrained form factors like notebooks and ultrathins.

 

[raghu78]

This is the future. Discrete boards, discrete ram, discrete SSD, are all going to go away.

Today we do not see the discrete sound card, the discrete network card, the discrete northbridge and southbridge in consumer PCs. I do not say the discrete GPU will go away. But its overall unit volume will shrink and prices will increase as process nodes get costlier and APUs eat away the entry level discrete GPU market.

 

[ShintaiDK]

Today we do not see the discrete sound card, the discrete network card, the discrete northbridge and southbridge in consumer PCs. I do not say the discrete GPU will go away. But its overall unit volume will shrink and prices will increase as process nodes get costlier and APUs eat away the entry level discrete GPU market.

We sure still see discrete NICs, discrete SBs and discrete NBs(AM3+) and a sound codec chip on the mobo.

You can find a few chips with most of it integrated like the Xeon D. But its certainly not the norm.

So lets stick to reality even tho it may seem incredible hard at the moment.

Integrated IGP volume also shrinks. Not just discrete GPUs. It again raises the economic question. How many are going to buy this product? And another part for your AM4 socket is the height that an interposer adds. It will mean that all non interposer chips will have an IHS that's 1mm more than needed. They could just as well add some GDDR5(x) sideport memory or cheaper EDRAM until the day when HBM may be able to reach a price level that will make it interesting in more than the flagship segment.

 

[raghu78]

We sure still see discrete NICs, discrete SBs and discrete NBs(AM3+) and a sound codec chip on the mobo.

You can find a few chips with most of it integrated like the Xeon D. But its certainly not the norm.

So lets stick to reality even tho it may seem incredible hard at the moment.

The motherboard has integrated a lot of components as has the CPU. Integration has been the consistent factor over the last 2 decades. HBM2 is a good example of that trend (where it integrates memory onto the CPU through an interposer) and could completely eliminate DDR4 memory.

Integrated IGP volume also shrinks. Not just discrete GPUs. It again raises the economic question. How many are going to buy this product? And another part for your AM4 socket is the height that an interposer adds. It will mean that all non interposer chips will have an IHS that's 1mm more than needed. They could just as well add some GDDR5(x) sideport memory or cheaper EDRAM until the day when HBM may be able to reach a price level that will make it interesting in more than the flagship segment.

I have a different view and that HBM2 is an ideal candidate for replacing system memory. We have to see how the future plays out. The idea of a PC on an interposer is exciting and has very interesting implications for form factor, board complexity, power efficiency and cost.

 

[JDG1980]

OEMs won't buy it unless it's cheap. Seriously.

Apple will buy it if it enables equivalent or better cost/perf than Intel CPU + low-midrange AMD/Nvidia dGPU in a smaller form factor and thermal envelope.

 

[JDG1980]

We sure still see discrete NICs, discrete SBs and discrete NBs(AM3+) and a sound codec chip on the mobo.

You can find a few chips with most of it integrated like the Xeon D. But its certainly not the norm.

It's not the norm just yet but it's starting to become the norm. And citing AM3, seriously? That dinosaur hasn't been updated since 2012. Modern platforms are becoming more and more SoC-like.

 

[el etro]

I think Polaris series is all about performance/$/mm², is something aimed to bring to mainstream price points today's enthusiast/halo performance. Is something like HD4870(P10) and HD5770(P11) again.

Vega will wipe our pockets later, bring all the performance enthusiasts are waiting for.