[TT] Pascal rumored to use GDDR5X..

[iiiankiii]

No, what you don't understand is few people care about any of these things except performance when it comes to high end video cards. Less power? Really? You think owners of exotic sports cars are arguing about who gets better fuel economy? So long as it isn't like 5 MPG, no one shopping them is going to care how "economical" they are. If the difference in power bills between a Nano and a 980ti is something you lose sleep over, you're involved in the wrong hobby. The only reason anyone cared about the terrible power usage of the 290x was because of the godawful OEM cooler AMD slapped on it which made the card miserable to live with.

Card size? Again who cares? Cases are probably one of the least frequently upgraded components of a PC. Has anyone been sitting around hoping for a Nano sized card because their case can't accommodate a regular sized video card and they have been stuck with onboard video for the last 3 years? I own a Silverstone Fotress ft02 pictured below (not my system)

As you can see, there is ample space for expansion cards. I have no complaints about this case and plan to keep it for years to come. Give me a non-idiotic reason why I should want a smaller sized video card when I have such a case?

As I've already discussed earlier in the thread I'm not blinded by shiny new things. The age of the tech means nothing to me.

Just because you can come up with some niche scenario where any of these things matter, doesn't mean the rest of us need to care. This isn't an Nvidia vs AMD debate. As long as the video card is within reasonable bounds of all the things you listed, the only one I care at all about is the performance. Up to this point HBM has not demonstrated any real world benefit. The speed of the individual components is meaningless to me, only the performance of the finished product. If Matrox comes out of no where and releases a card with EDO DRAM that is faster than a card using HBM, that's what I will buy.

I agree. Power consumption and size goes out the window with high end parts. However, HBM is faster than GDDR5. The potential is there.

More importantly, you're overlooking an important part about HBM. It consumes less power. Now, I know. I just said it doesn't matter at the high end. But, it does matter. By consuming less power, NV/AMD can use that extra TDP on performance. NV/AMD seems to recognize that the max TDP for a reasonably sized high end GPU is around 250-275 watts.

Anandtech on GTX 980TI clock speed increase and performance vs power consumption.

The gains from this overclock are a very consistent across all 5 of our sample games at 4K, with the average performance increase being 20%. ......

..... The cost of that 20% overclock in terms of power and noise is similarly straightforward. You're looking at an increased power cost of 30W or so at the wall – in-line with the 25W increase in the card’s TDP

By using HBM, you can expect to save roughly 20 watts or so in TDP.

What’s the real-world advantage of a 15-20W reduction in DRAM power consumption? Besides being able to invest that in reducing overall video card power consumption, the other option is to invest it in increasing clockspeeds.

Now, I understand that 20 watts reduction doesn't sound like much. Fortunately, that 20 watts translates to 10-15% in performance increase on a GTX 980TI. So, it does matter.

 

[Pariah]

Valid points. But again, at least for myself, I only care about the end results. I don't care about the methods used to get there. HBM is going to be the future, but that doesn't mean the future is now. HBM may well allow for a 20W reduction in power which can be used for better performance, but that doesn't mean that a card using HBM is guaranteed to be faster than a card using some other memory. I'm not concerned with what memory a video card has. All else being relatively equal, I will by the fastest card for the games I play at the setting I play in my budget regardless of the memory it uses.

 

[ShintaiDK]

Seems the nano is even shorter than the 970 so somewhat true. Most definitely cannot physically have smaller gddr5 cards because you need space for the power and VRAM.

The Nano is shorter than ITX boards. Meaning it could just as well be an inch longer without losing anything. The next step is low profile, further reductions from the MiniITX length is just foolish.

 

[ShintaiDK]

Instead of a GTX 970 card class card in an "mitx" form factor, we can get a GTX 980Ti class card with HBM in an "mitx" form factor, except it's in that form factor because it's just that efficient.

Don't see why any person in the world would not be excited for that, but apparently you are not.

While HBM is the future and its great on a lot of fronts. I dont see why a 980TI Mini with the same throttle/clock reductions as the Nano couldn't happen.

Looks to be space enough.

The Nano became a bit silly when it got smaller than the MiniITX boards its supposed to sit in. Its like a competition who is the shortest midget.

What HBM however can do is to bring great performance to future low profile cards.

 

[ShintaiDK]

GDDR5X is a stop-gap solution that is logical and reasonable while we wait the full maturation of HBM2 (and further generations).

^^ Perfectly summoned up.

Look at HBM1. It serves 3 tiny volume gaming products from the absolute smallest vendor. And even that seems to be hard to manufactor in the required volume.

 

[dacostafilipe]

Look at HBM1. It serves 3 tiny volume gaming products from the absolute smallest vendor. And even that seems to be hard to manufactor in the required volume.

So did GDDR5 when released on the 4870.

 

[Keysplayr]

That was exactly my point Keysplayer actually. It's my own statement everyone is responding to, so yes, that's what I said. That HBM has a great benefit in reducing card size.

Why ANYONE would be upset with this is beyond me.


It allows for a smaller and more efficient card to be made. That's why I'm excited for HBM. Why you guys want to bring the Nano into this, I don't know. HBM allows for smaller designs as shown by Nvidia's own debut of Pascal. So yes, why would I NOT be excited for that? Instead of a full length GPU needed, we can get big pascal performance that can fit into a mini case? Or Arctic Islands?

Being able to fit more powerful GPUs into smaller spaces is cool... and exciting.

Are you saying you're against that and would prefer Pascal to be as long of a card as possible?

No. But this seems to be the automatic interpretation when somebody is complaining about the "argument" instead of the actual product.

In other words, it's nice to see things get smaller instead of bigger. Certainly. I think it's a positive. But for the love of all that's holy, don't pimp it as an end all be all selling point that can ONLY be done by a vendor who will (reportedly) have the only access to HBM/2 anywhere in the near future. Not saying you directly, but that is the ever tiresome mantra vibe that never ceases.
Like a machine, some of these folks are.

Why Nano? Isn't that the first small form factor high end GPU that is able to achieve it's size due to HBM memory? Or so touted? Why else would it be mentioned? Because people began bashing future cards with GDDR5X because they "won't be small" and more costly.

The SFF crowd is a small crowd, but by the talk in here you'd think it was everyone who was in this crowd.

 

[ShintaiDK]

So did GDDR5 when released on the 4870.

But it didn't require 2 additional steps in the form of TSV and interposer. Its always easier to ramp production of something that is essentially more of the same. While its very hard to produce something entirely new, and even more importantly, bring prices down.

 

[AtenRa]

While HBM is the future and its great on a lot of fronts. I dont see why a 980TI Mini with the same throttle/clock reductions as the Nano couldn't happen.

Looks to be space enough.

1. GM200 VRM may not be possible at this small PCB size.
2. If you lower the TDP it will lοose more performance than Nano did vs the Fury because of the GDDR-5.
3. GM200 has a 384bit memory controller and you will need 6GB of GDDR-5, this will take additional space.

 

[ShintaiDK]

1. GM200 VRM may not be possible at this small PCB size.
2. If you lower the TDP it will lοose more performance than Nano did vs the Fury because of the GDDR-5.
3. GM200 has a 384bit memory controller and you will need 6GB of GDDR-5, this will take additional space.

1 and 3 I think is possible within the MiniITX size. Specially since the GTX970 doesn't seem to have the same amount of components on the backside in the same way like the Nano.

For 2, do you have any numbers to back up this claim?

 

[guskline]

I read this forum post which lays it out in terms I understand.
Pavan Biliyar · San Diego State University

To see the benefit of higher bandwidth, we only need an example of two identical GPUs where each relies on different memory bandwidth options. Such a case exists, ironically, with Intel IGP. The Crystalwell HD5100 and HD5200 share the same 40-eu GPU but only HD5200 has the eDRAM with triple the bandwidth as DDR3-1600. The difference is a 70% increase in 3DMark performance. Mathematically, the square root of 3 is 1.73, showing that an exponential level of bandwidth is needed to gain any reasonable performance from a weak GPU. Quadrupling the bandwidth would only double performance, etc.

Stacked VRam, HBM and eDRAM are just methods to the same result to ncrease memory bandwidth.

AMD's GPUs are weak compared to nVidia, the only chance they have to come close is to raise the bandwith. Tahiti only came within percents of GK104 because it had a 384-bit bus against Kepler's 256-bit bus. But once the first Titan raised to 384-bit, AMD needed 512-bit on Hawaii. But once nVidia raised the clocks on the GDDR5, AMD couldn't just increase the interface again, it would have raised the power rating too much: Enter HBM.

I suspect, given how close HBM 1.0 is to nVidia's 384-bit GGDR5 at 7GHz (a difference of just 45%), if nVidia managed to raise the GDDR5 clock to 9GHz, the advantage would shallow out:

4096-bit at 1GHz effective, divide by 8-bits per byte = 512GB/s
384-bit at 9GHz effective, divide by 8-bits per byte = 432GB/s

We see it with Intel IGP, the Skylake IGP in i7-6700K is the sane 40-eu as Crystallwell, but relies on a faster DDR4-2133+ shared memory bandwidth, which closes the gap between HD5100 and HD5200.

Overall benefit is up to the user, it isn't a general one-size-fits-all obvious. If you game in such a way that doesn't need the bandwidth (nVIdia has astong GPU paired with weak bandwidth, and their gains are in lower resolutions of 1080p and 1440p), then you will have no need of HBM until more powerful graphics cards come.

 

[littleg]

I read this forum post which lays it out in terms I understand.
Pavan Biliyar · San Diego State University

To see the benefit of higher bandwidth, we only need an example of two identical GPUs where each relies on different memory bandwidth options. Such a case exists, ironically, with Intel IGP. The Crystalwell HD5100 and HD5200 share the same 40-eu GPU but only HD5200 has the eDRAM with triple the bandwidth as DDR3-1600. The difference is a 70% increase in 3DMark performance. Mathematically, the square root of 3 is 1.73, showing that an exponential level of bandwidth is needed to gain any reasonable performance from a weak GPU. Quadrupling the bandwidth would only double performance, etc.

Stacked VRam, HBM and eDRAM are just methods to the same result to ncrease memory bandwidth.

AMD's GPUs are weak compared to nVidia, the only chance they have to come close is to raise the bandwith. Tahiti only came within percents of GK104 because it had a 384-bit bus against Kepler's 256-bit bus. But once the first Titan raised to 384-bit, AMD needed 512-bit on Hawaii. But once nVidia raised the clocks on the GDDR5, AMD couldn't just increase the interface again, it would have raised the power rating too much: Enter HBM.

I suspect, given how close HBM 1.0 is to nVidia's 384-bit GGDR5 at 7GHz (a difference of just 45%), if nVidia managed to raise the GDDR5 clock to 9GHz, the advantage would shallow out:

4096-bit at 1GHz effective, divide by 8-bits per byte = 512GB/s
384-bit at 9GHz effective, divide by 8-bits per byte = 432GB/s

We see it with Intel IGP, the Skylake IGP in i7-6700K is the sane 40-eu as Crystallwell, but relies on a faster DDR4-2133+ shared memory bandwidth, which closes the gap between HD5100 and HD5200.

Overall benefit is up to the user, it isn't a general one-size-fits-all obvious. If you game in such a way that doesn't need the bandwidth (nVIdia has astong GPU paired with weak bandwidth, and their gains are in lower resolutions of 1080p and 1440p), then you will have no need of HBM until more powerful graphics cards come.

'Just' 45%

Bandwidth is only useful as long as the core can use it anyway. If your core can only handle 200GB/s then you can throw 500, 1000, 10,000 GB/s bandwidth at it and you won't see much of a gain.

 

[AtenRa]

For 2, do you have any numbers to back up this claim?

GTX-980Ti has 6GB GDDR-5 at 7012MHz, that will be close to 50-60W of power for the memory alone.

HBM has ~40% less power consumption than GDDR-5

Fury Nano has 4GB of HBM, that will be close to 20-25W of power.


So the more you lower the power consumption of the Graphics Card, the more you have to give to the GPU if you use HBM, because Memory power consumption will remain constant.

 

[ShintaiDK]

GTX-980Ti has 6GB GDDR-5 at 7012MHz, that will be close to 50-60W of power for the memory alone.

HBM has ~40% less power consumption than GDDR-5

Fury Nano has 4GB of HBM, that will be close to 20-25W of power.


So the more you lower the power consumption of the Graphics Card, the more you have to give to the GPU if you use HBM, because Memory power consumption will remain constant.

So you dont have any numbers.

And I thought this had already been posted.
http://www.anandtech.com/show/9266/amd-hbm-deep-dive/4

 

[Keysplayr]

GTX-980Ti has 6GB GDDR-5 at 7012MHz, that will be close to 50-60W of power for the memory alone.

HBM has ~40% less power consumption than GDDR-5

Fury Nano has 4GB of HBM, that will be close to 20-25W of power.


So the more you lower the power consumption of the Graphics Card, the more you have to give to the GPU if you use HBM, because Memory power consumption will remain constant.

Sigh..... A TitanX with 12GB of 7Gbps eDRAM consumes 31.5W of power.

At one third (1/3rd) of the amount of memory, 4GB of HBM uses 14.6W of power.

Where are you getting your numbers from, AtenRa?

 

[Erenhardt]

Glad that memory controllers are powered by fanboy hate. Otherwise those too would bite into power budget.

Anyway, we arrive at half the power for two time the bandwidth just looking at memory modules itself. Quite astonishing and something worth pursuing, not a measly 20% perf/watt that is advertised by nv all over the webs, or better:

 

[AtenRa]

Sigh..... A TitanX with 12GB of 7Gbps eDRAM consumes 31.5W of power.

At one third (1/3rd) of the amount of memory, 4GB of HBM uses 14.6W of power.

Where are you getting your numbers from, AtenRa?

Yea my bad forgot to say that the numbers were made for the comparison. But they are very close to the real thing, almost half the power than the GDDR-5 but with higher bandwidth.
So with the HBM you have more power to give to the GPU as you lower the power consumption of the Card and that means you loose less performance.

 

[PPB]

The Nano became a bit silly when it got smaller than the MiniITX boards its supposed to sit in. Its like a competition who is the shortest midget.

What HBM however can do is to bring great performance to future low profile cards.

It is not when the PCI-E Power connector is on the back. This saves width in the case (you can basically have your GPU PCB width be the limiting factor for the overall width of the case) and makes you have as tiny as 17-17.5cm depth in your case. You can even add a slim fan in that space without exceding the 17.5cm depth footprint.

The only thing silly here is your argument.

 

[tential]

You know better than this Tential - or at least I hope so.

Nobody has said that GDDR5 is a better technology than HBM, but you provide no context for your rant. What everyone has been saying, at least those of us who think this is a logical step, is that HBM just isn't ready for primetime yet on a broad mainstream basis.

Silverforce said as much when he said:



Exactly. GDDR5X is a stop-gap solution that is logical and reasonable while we wait the full maturation of HBM2 (and further generations). This isn't inconsistent.

For you to try to paint people as anti-HBM because we're shills for NV or whatever is laughable. I don't think anyone can accuse Silverforce of being an NV shill, if you look at his post history, but at least he's intellectually honest enough to concede the current limitations of HBM without descending into a rant attacking everyone who understands this as unpaid shills for NV.

Honestly, why is it so difficult to understand this?

I made no rant.
All I said was hbm brings an overall reduction in card size. That's the only comment I made. People then decided to dispute that.

I never made any comment on whether it needs to be used now (I actually said I have 0 care which platform is used. I don't care, I just want performance.).

I simply said hbm brings an overall reduction in card size.

What everyone inferred after that I don't understand.

 

[Azix]

The Nano is shorter than ITX boards. Meaning it could just as well be an inch longer without losing anything. The next step is low profile, further reductions from the MiniITX length is just foolish.

For some or most people sure. But it still means it can be used in a wider variety of builds (i.e. more compact prebuilt systems or space for more components in a compact build)

Sigh..... A TitanX with 12GB of 7Gbps eDRAM consumes 31.5W of power.

At one third (1/3rd) of the amount of memory, 4GB of HBM uses 14.6W of power.

Where are you getting your numbers from, AtenRa?

this 31.5 W figure was just a guess that does not make sense. They were saying a titan x with higher clocked x3 VRAM only used 1.5W more than a 290x with just 4GB VRAM clocked lower, all because of its 512bit bus vs the titan X and its 384bit. makes no real sense.

 

[tential]

No. But this seems to be the automatic interpretation when somebody is complaining about the "argument" instead of the actual product.

In other words, it's nice to see things get smaller instead of bigger. Certainly. I think it's a positive. But for the love of all that's holy, don't pimp it as an end all be all selling point that can ONLY be done by a vendor who will (reportedly) have the only access to HBM/2 anywhere in the near future. Not saying you directly, but that is the ever tiresome mantra vibe that never ceases.
Like a machine, some of these folks are.

Why Nano? Isn't that the first small form factor high end GPU that is able to achieve it's size due to HBM memory? Or so touted? Why else would it be mentioned? Because people began bashing future cards with GDDR5X because they "won't be small" and more costly.

The SFF crowd is a small crowd, but by the talk in here you'd think it was everyone who was in this crowd.

Well my comment was simple. It was that hbm allows for smaller cards.

That's it. What you guys decided to run with and fight over I don't get. I'm not saying that new cards without hbm suck or are undesirable. I've even said I will buy the best card for me. Hbm is not a selling point to me at all.

Just that hbm allows for smaller cards. It's such a simple statement I don't get why you'd want to take it further to suggest a card that is available today or in the near future will suck or will be great.

That's why I have specifically tried to avoid mentioning any real cards because I am simply saying it's nice hbm is coming and will make it easier for cards to be smaller.

Gddr5x isn't terrible or anything. It's a great stopgap between hbm and gddr5.....

I think I know the posts / posters you're talking about and my own comments have nothing to do with those comments, which were bashing Gddr5x and saying hbm is better etc etc etc.

I haven't made any claim Gddr5x is terrible, worse, etc. Either. Nothing against it.

 

[Techhog]

Sigh..... A TitanX with 12GB of 7Gbps eDRAM consumes 31.5W of power.

At one third (1/3rd) of the amount of memory, 4GB of HBM uses 14.6W of power.

Where are you getting your numbers from, AtenRa?

That would imply that HBM is less efficient than GDDR5, wouldn't it? I think you need to source your numbers as well...

 

[96Firebird]

That would imply that HBM is less efficient than GDDR5, wouldn't it? I think you need to source your numbers as well...

The source for the numbers is in the post above Keysplayer's.

 

[littleg]

Never mind

 

[Keysplayr]

Yea my bad forgot to say that the numbers were made for the comparison. But they are very close to the real thing, almost half the power than the GDDR-5 but with higher bandwidth.
So with the HBM you have more power to give to the GPU as you lower the power consumption of the Card and that means you loose less performance.

Yeah. Half the power with 1/3rd the memory. Fascinating.