[TT] Pascal rumored to use GDDR5X..

[Headfoot]

So, it's confirmed.

http://semiaccurate.com/forums/showthread.php?t=8757

Im more interested in that this guy is claiming sept/oct for Pascal release. Another year on 28nm unless AMD pulls a rabbit out of the hat

 

[Techhog]

How much of that is due to new to the Fury arch and how much is due to HBM1? You're attributing all the gains to HBM1 and I wonder how you can source those claims. Let's see.

What changes? There were barely any changes at all. Based on 285 vs 280, GCN 1.2 saved, at most, 20W over GCN 1.1, and that's high-balling it.

https://www.techpowerup.com/reviews/AMD/R9_Nano/31.html

nVidia did better with GDDR5.
Yes, HBM allowed AMD to build a card which is faster and uses less power but they needed HBM for this.

nVidia has changed the architecture from Kepler and has released a whole family of chips from low-end to high-end.

nVidia is faster with GDDR, uses less power and has more memory.

I knew that someone was going to say this. Yes, of course a new architecture does more than changing the memory, but what's you point? That Maxwell would have been a worse architecture if they had used HBM? Because if that's not your point, your entire post wasn't saying anything even remotely relevant. Or do you think that HBM is the sole reason that AMD couldn't have a successor to GCN ready this year? I'm talking about the relative impact of HBM. You're adding hundreds of variables for no reason.

 

[sontin]

AMD needed HBM to have the room for more transistors without running out of power...

Without HBM Fury X would use 350W+ instead of <=300W.

 

[AtenRa]

AMD needed HBM to have the room for more transistors without running out of power...

Without HBM Fury X would use 350W+ instead of <=300W.

So you agree that the same Pascal chip with HBM will be much much better than GDDR5X.

 

[Techhog]

AMD needed HBM to have the room for more transistors without running out of power...

Without HBM Fury X would use 350W+ instead of <=300W.

So, you agree that HBM makes a huge difference in power consumption, and thus agree with my point?

 

[Silverforce11]

AMD needed HBM to have the room for more transistors without running out of power...

Without HBM Fury X would use 350W+ instead of <=300W.

Yup. The power savings isn't just from the memory chips, but a big bus GDDR5 memory controller takes a large chunk of die space and TDP.

One don't have to imagine much to see how such a huge benefit could come in very handy for Teslas (IIRC, next year is Intel's next-gen Phi HPC GPU's debut?) as well as high-end GP200 SKUs, more TDP available at play for performance = good.

 

[ShintaiDK]

This is better for the mid range cards that would never have got HBM anyway as GDDR5X is better then GDDR5. I suppose the only question is what do the GTX 970/980 class cards get - HBM or GDDR5X?

I would guess GDDR5X. The cost structure of HBM makes it hard to justify in 300-400$ cards until the technology not only matures a lot more. But also get much higher production, including multiple producers.

 

[dacostafilipe]

Will Pascal mid-range cards be that fast that today's GDDR5 wouldn't be enough? I don't think so.

HBM (1 and 2) just makes sense for the future in terms of performance, costs and features.

 

[ShintaiDK]

Will Pascal mid-range cards be that fast that today's GDDR5 wouldn't be enough? I don't think so.

HBM (1 and 2) just makes sense for the future in terms of performance, costs and features.

GTX970 and GTX980 successors? Absolutely. They need GDDR5X or HBM unless you maybe pair them with a 384/512bit bus.

 

[AnandThenMan]

The cost structure of HBM makes it hard to justify in 300-400$ cards...

What is the cost structure of HBM versus GDDR memory.

 

[ShintaiDK]

What is the cost structure of HBM versus GDDR memory.

Limited supply, premature, currently 1 supplier, TSV, Interposer etc.

It shouldn't be a surprise. Is DDR4 cheaper than DDR3 today?

 

[AnandThenMan]

Limited supply, premature, currently 1 supplier, TSV, Interposer etc.

It shouldn't be a surprise. Is DDR4 cheaper than DDR3 today?

You didn't answer the question and provided no data just guesses and your opinion.

 

[ShintaiDK]

You didn't answer the question and provided no data just guesses and your opinion.

Right I forgot. HBM is completely different from anything else in the industry. Super cheap, easy to implement, massive supply from multiple producers and everything is a piece of cake and a home run from day 1.

Better? Because that's what you want to hear isn't it.

The interposer alone is 5-10$ over GDDR5(X). Something that will never change. Its just a new baseline extra cost. TSV implementation isn't free either. And then we can talk about the supply and production of the chips. Even if everything goes smooth, HBM as chips alone may not even reach GDDR5(X) price parity before a few years. And then there is still the TSV and interposer.

GDDR5X didn't come out of the blue because everything is only rosy in HBM land. Its a hotfix until HBM can deliver.

 

[Gikaseixas]

Right I forgot

A simple and honest "i don't know" would be enough and save us from your usual biased lines

 

[Gikaseixas]

What is the cost structure of HBM versus GDDR memory.

Don't know but i would guess that HBM is more expensive and HBM2 would add to that. GDDR5X would end up costing a bit less therefore could be suited to mainstream parts. Just my 2 cents

 

[AtenRa]

That is only the cost of the HBM modules,

Due to HBM you also have,

Lower power consumption and lower thermals = less copper = Lower BOM
Smaller Graphics Card sizes = smaller PCBs = Lower BOM
Smaller Graphics Card boxes = Lower packaging cost

Best oriented for Laptops = higher margins = higher profits

Even if the overall cost of the final product with an HBM is 10-20% higher than the GDDR5X, you will still have a product that every OEM would like to install in their Laptops because of the smaller size, lower thermals, higher performance than those GDDR5X products.

HBM may not be ok for the sub $100-150 Desktop market in 2016 but im betting that it will be the next big thing for the mid-high GPU Laptop market where margins are higher and you can absorb the extra cost.

 

[littleg]

If that's the case then they'll surely go with HBM for their mid - high end laptop GPUs, right? We've so far only heard about GDDRX5 going in lower end cards where it kind of makes sense (for now anyway).

 

[maddie]

Right I forgot. HBM is completely different from anything else in the industry. Super cheap, easy to implement, massive supply from multiple producers and everything is a piece of cake and a home run from day 1.

Better? Because that's what you want to hear isn't it.

The interposer alone is 5-10$ over GDDR5(X). Something that will never change. Its just a new baseline extra cost. TSV implementation isn't free either. And then we can talk about the supply and production of the chips. Even if everything goes smooth, HBM as chips alone may not even reach GDDR5(X) price parity before a few years. And then there is still the TSV and interposer.

GDDR5X didn't come out of the blue because everything is only rosy in HBM land. Its a hotfix until HBM can deliver.

I guess your universe doesn't accommodate the fact that this interposer allows at the very least, a simpler AND smaller PCB.
Yes, there are new costs.
Yes, there are new savings.
Net is the important figure.
You don't know and neither do most of us here, so please don't try to portray yourself as an expert. If you do get verifiable numbers, please post them, as I'm sure a lot of us would like an accurate view of costs.

 

[ShintaiDK]

PCB cost savings is pennies.

 

[AnandThenMan]

Better? Because that's what you want to hear isn't it..

No I wanted you to back up your claim, which you almost never do. And everyone would like to see what the cost structure of HBM actually is, the thing you claim to know all about. You specifically talked about cost structure.

The cost structure of HBM makes it hard to justify in 300-400$ cards

If you don't know what the total cost is (you obviously have no idea) then don't make definitive statements.

PCB cost savings is pennies.

How many pennies? And back up your claim.

 

[ShintaiDK]

When you need additional steps and components the cost structure goes up. And this is something HBM suffers from as a baseline cost. Limited supply and production problems just adds on top.

I can understand HBM is a tender subject for some. But patience and it will work out. GDDR5X is only here because Micron knows HBM is so cost ineffective currently and for a foreseeable future. That they can sell GDDR5X and recoup any ROI in the timeframe.

 

[AnandThenMan]

When you need additional steps and components the cost structure goes up. And this is something HBM suffers from as a baseline cost.

I can understand HBM is a tender subject for some.

HBM is the future of memory, the only one that doesn't seem to grasp this is you. It it likely HBM costs more overall versus GDDR5 but this is expected as is the case with almost every new bit of tech. Next you're going to tell us that no one should go with a new manufacturing node because initially it costs more.

 

[ShintaiDK]

HBM is the future of memory, the only one that doesn't seem to grasp this is you. It it likely HBM costs more overall versus GDDR5 but this is expected as is the case with almost every new bit of tech. Next you're going to tell us that no one should go with a new manufacturing node because initially it costs more.

No, I have even posted that GDDR5X is a hotfix.

HBM still looks like a premature tech that needs to evolve further to get a solid ground. While GDDR5X is a bit of a hotfix. It solves the current short term issues and provides something HBM cant do yet.

You dont see CPUs with HBM or HMC memory yet either do you? Its essentially all about cost. This will be fixed over time. But its a long process.

Native quadcores was also the future. Yet AMDs first native quadcore was a disaster due to cost. Everything got its time and place for mass usage.

 

[lilltesaito]

When you need additional steps and components the cost structure goes up. And this is something HBM suffers from as a baseline cost.

I can understand HBM is a tender subject for some.

No doubt that stuff cost money, people are trying to say that we dont know what this cost is and that it might not be as much when you start taking out the cost of the ddr components.

Since we do not have the cost we can not say either way if it is to much or not.

Edit: Also since the new cards most likely had HBM in mind, would it not also cost money to redo the cards to fit ddr instead?

 

[ShintaiDK]

Also since the new cards most likely had HBM in mind, would it not also cost money to redo the cards to fit ddr instead?

It depends how many cards have been developed for it. I dont think either AMD or NVidia did so for mainstream cards.