AMD HBM High Bandwidth Memory Technology Unveiled @ [H]

[FrgMstr]

AMD HBM High Bandwidth Memory Technology Unveiled - AMD unveils a new memory technology that we will see incorporated in its video cards that we will see at some point in the future. It is a true departure from our current GDDR ways of doing things and makes some huge promises as well. HBM promises around 4X the memory bandwidth in a smaller footprint than one GDDR5 chip.

 

[Trepidati0n]

Meh....this has been talked about for a long time by many companies (nearly a decade). However, graphics cards are probably one of thee best places for this tech though.

 

[Silentbob343]

Really looking forward to the size benefits. I had to trim part of the fan shroud to get my card to fit in to my MITX case. Didn't realize how much space is being taken up by the RAM chips on a modern GPU.

Wonder what kind heat reduction since they use less power.

 

[FrgMstr]

Really looking forward to the size benefits. I had to trim part of the fan shroud to get my card to fit in to my MITX case. Didn't realize how much space is being taken up by the RAM chips on a modern GPU.

Wonder what kind heat reduction since they use less power.

Well, going by the slides, the HBM stacks are using a reduced voltage at 1.3v compared to GDDR5's 1.5, and AMD is claiming ~35GB/s of bandwidth per watt, compared to GDDR5's 10.66GB/s per watt.

 

[FrgMstr]

Meh....this has been talked about for a long time by many companies (nearly a decade). However, graphics cards are probably one of thee best places for this tech though.

You are correct, die stacking has been talked about for a long time, and there is a reason for this.....it brings with it some great advantages.

 

[Zarathustra[H]]

HBM is undoubtedly the future memory tech, at the very least for Video cards.

I question how much of a practical performance benefit it will actually have in this generation of video cards - however.

Sure, it may MASSIVELY increase the available video ram bandwidth, but we already know from overclocking that current gen GPU's only marginally benefit from upping the memory bandwidth, and there is no reason to believe that this upcoming generation of GPU's from AMD will be any different.

Power savings have also been listed as a huge benefit of HBM. And sure, the 3x better performance per watt sure is impressive. Considering - however - how small a portion of the overall video card power use is in the RAM, this will have a rather marginal impact on high performance products. (it will be huge on low power mobile devices though)

Lets say you have a 300W video card with GDDR5 RAM, which uses 275W for the GPU and 25 W for the ram. Cut the RAM power use by a factor of 3, and you are now using 8.3W for the RAM, giving you an additional 16.7W to use for the GPU while still staying within the 300W power envelope.

So that's moving from 275w to 291.7W, and increase of ~6%. Not bad. Every little bit counts, but it's not enough to be a game changer

If anything, limited early supply of early HBM production will be the biggest impact of going HBM this gen, in reducing the availability of these cards.

I hate to say it, but Nvidia's "wait and see" approach, going with HBM in its second generation instead, was likely the smarter approach, though I will be happy to be proven wrong on launch day

Over time HBM will be of huge importance, but its importance will only grow slowly, bit by bit each generation. Those expecing an overnight change because OMG HBM are going to be hugely disappointed.

 

[THUMPer]

The first gen of HBM is already leaps and bounds ahead of GDDR5. Can't wait.

 

[Armenius]

HBM is undoubtedly the future memory tech, at the very least for Video cards.

I question how much of a practical performance benefit it will actually have in this generation of video cards - however.

Sure, it may MASSIVELY increase the available video ram bandwidth, but we already know from overclocking that current gen GPU's only marginally benefit from upping the memory bandwidth, and there is no reason to believe that this upcoming generation of GPU's from AMD will be any different.

Power savings have also been listed as a huge benefit of HBM. And sure, the 3x better performance per watt sure is impressive. Considering - however - how small a portion of the overall video card power use is in the RAM, this will have a rather marginal impact on high performance products. (it will be huge on low power mobile devices though)

Lets say you have a 300W video card with GDDR5 RAM, which uses 275W for the GPU and 25 W for the ram. Cut the RAM power use by a factor of 3, and you are now using 8.3W for the RAM, giving you an additional 16.7W to use for the GPU while still staying within the 300W power envelope.

So that's moving from 275w to 291.7W, and increase of ~6%. Not bad. Every little bit counts, but it's not enough to be a game changer

If anything, limited early supply of early HBM production will be the biggest impact of going HBM this gen, in reducing the availability of these cards.

I hate to say it, but Nvidia's "wait and see" approach, going with HBM in its second generation instead, was likely the smarter approach, though I will be happy to be proven wrong on launch day

Over time HBM will be of huge importance, but its importance will only grow slowly, bit by bit each generation. Those expecing an overnight change because OMG HBM are going to be hugely disappointed.

Someone always has to be first, and I never lose my respect for AMD usually being the first to dip their toes in the water of new tech for video cards. There are still a lot of questions as to what practical benefit HBM will provide as far as real-world gaming performance is concerned, and thankfully we will soon find out with the 300-series launching next month. But one still has to wonder what is in store with PCI-E 4.0 and NVLink on the horizon.

 

[Silentbob343]

HBM is undoubtedly the future memory tech, at the very least for Video cards.

I question how much of a practical performance benefit it will actually have in this generation of video cards - however.

Hell, I'll be happy for the reduction in package size alone. There are some great m-itx cases around and modern GPUs are a bit too long for the majority of them. If they can make a card in the 200mm range it opens up case options. There are already niche m-itx GPUs, but they are more expensive compared to their std length length siblings.

 

[Nauseous]

looks cool cant wait for reviews of the card.

 

[Dawill]

More bandwidth is always good. However I'm more concerned about what arstechnica says about this. Their article (linked) says that 4gb max has been confirmed?? That would mean the 390x is limited to competing with the 980 and isn't any competition for the titan-x. I'm a new owner of a titan-x, but I was still hoping for AMD to show up for some competition as it's good for the market. 4GB max would be very bad IMO.

http://arstechnica.com/information-...nfirms-4gb-limit-for-first-hbm-graphics-card/

 

[defaultluser]

Hell, I'll be happy for the reduction in package size alone. There are some great m-itx cases around and modern GPUs are a bit too long for the majority of them. If they can make a card in the 200mm range it opens up case options. There are already niche m-itx GPUs, but they are more expensive compared to their std length length siblings.

The majority of that size in these smaller mitx cards is for the heatsink and cooler, and power supply (VRMs). The memory bus on the GTX 970 is only 256-bit, and there are only 8 GDDR5 devices around the core.

Witness the reference card design on the GTX 750 Ti, where they have a tiny cooler, but the memory doesn't take up much board space outside of the heatsink area required to cool the chip:

http://www.techpowerup.com/reviews/NVIDIA/GeForce_GTX_750_Ti/4.html

And then the mitx GTX 970 boards uses 3x the power, and takes up most of the increased board space with power supply. The card is an inch longer, but not because of the 4 extra memory dies:

http://www.bit-tech.net/hardware/graphics/2015/01/15/asus-geforce-gtx-970-directcu-mini-review/1

Other differences: the GTX 750 Ti heatsink is a simple aluminum with copper core affair, while the 970 DirectCU Mini (contrary to the name) actually uses a vapor chamber, which is much more expensive. You don't go to that expense unless the heat density is enormous, and you can't make the cooling solution any larger.

Nvidia has hit the point where, Titan aside, power is the limiting factor to card size, and not memory chips. Removing 150w from a 7 inch card using air cooling only is a difficult thing to do especially when you consider noise.

Now, on notebooks this could be more impressive, but once again the memory is less than half of the MXM card size.

 

[Zarathustra[H]]

More bandwidth is always good. However I'm more concerned about what arstechnica says about this. Their article (linked) says that 4gb max has been confirmed?? That would mean the 390x is limited to competing with the 980 and isn't any competition for the titan-x. I'm a new owner of a titan-x, but I was still hoping for AMD to show up for some competition as it's good for the market. 4GB max would be very bad IMO.

http://arstechnica.com/information-...nfirms-4gb-limit-for-first-hbm-graphics-card/

Hoping this isn't the case, but it wouldn't surprise me considering all the reports I've heard of HBM shortages.

 

[tungt88]

Hoping this isn't the case, but it wouldn't surprise me considering all the reports I've heard of HBM shortages.

"An obvious concern is the limit of 4GB of memory for the upcoming Fiji GPU – even though AMD didn’t verify that claim for the upcoming release, implementation of HBM today guarantees that will be the case. Is this enough for a high end GPU? After all, both AMD and NVIDIA have been crusading for larger and larger memory capacities including AMD’s 8GB R9 290X offerings released last year. Will gaming suffer on the high end with only 4GB? Macri doesn’t believe so; mainly because of a renewed interest in optimizing frame buffer utilization. Macri admitted that in the past very little effort was put into measuring and improving the utilization of the graphics memory system, calling it “exceedingly poor.” The solution was to just add more memory – it was easy to do and relatively cheap. With HBM that isn’t the case as there is a ceiling of what can be offered this generation. Macri told us that with just a couple of engineers it was easy to find ways to improve utilization and he believes that modern resolutions and gaming engines will not suffer at all from a 4GB graphics memory limit. It will require some finesse from the marketing folks at AMD though ..."

from: PcPer

So yeah, probably 4GB (2x4GB on the dual GPU, o'course). Let's see how this new tech holds up in RL, though, on the high-rez, memory-hungry games. Can't wait for the HardOCP review

 

[Armenius]

Hoping this isn't the case, but it wouldn't surprise me considering all the reports I've heard of HBM shortages.

According to the article the limit of 4GB may be due to the physical size of the core and the limit of HBM1 to stacks that are 4 deep.

 

[Elf_Boy]

I can see the use in a GPU.

I am wondering what it would do if properly set up on my motherboard (well a future board).

Is it not true that one of the big bottle necks on a main board is waiting for dram? If I remember right lvl 2/3 caching takes up a bit of space on the CPU die and if that could be moved off the CPU wouldn't it seriously decrease heat and allow for higher clocks?

 

[cageymaru]

I can see the use in a GPU.

I am wondering what it would do if properly set up on my motherboard (well a future board).

Is it not true that one of the big bottle necks on a main board is waiting for dram? If I remember right lvl 2/3 caching takes up a bit of space on the CPU die and if that could be moved off the CPU wouldn't it seriously decrease heat and allow for higher clocks?

What if they shrunk the CPU to 14nm using FinFET technology and made the die just as big or bigger than today? That way you could use HBM on the die, generate less heat because of it, and keep it all in one package. That would be pretty cool also.

 

[Brackle]

Well if 2 980 GTX's are fine for 4k why wouldn't a 390x? (when it comes to memory).

Sure there are games that might need more then 4GB of memory, but you will run out of GPU horsepower quicker then hitting a Vram limit at 4k.

But yes I think 4GB is a lil low for future proofing. I think Nvidia will release a 980TI with 6GB of memory and people will buy it because it has 2GB more memory.

Yet 2 980's in SLI are good for 4k gaming.

 

[Chris_B]

Are there any real cons to stacked memory? The only thing i can think of is more power\heat being focused in a smaller area, with standard gddr it was more spread out over a larger area. Still seems odd that the rumours are pointing to these new cards being substantially smaller than what we've had previously. I think most people by now had got used to having to apply for planning permission to install a new card

 

[TheOneKnownAsMe]

Are there any real cons to stacked memory? The only thing i can think of is more power\heat being focused in a smaller area, with standard gddr it was more spread out over a larger area.

Yeah, basically the biggest "con" would seem to be heat. It's all packed into a much smaller, tighter area.

 

[Brackle]

Are there any real cons to stacked memory? The only thing i can think of is more power\heat being focused in a smaller area, with standard gddr it was more spread out over a larger area. Still seems odd that the rumours are pointing to these new cards being substantially smaller than what we've had previously. I think most people by now had got used to having to apply for planning permission to install a new card

Possibly? Maybe overclocking the memory might be harder to do? The only con to me is the max memory that can be stacked. This will be a non-issue in the future though.

No one really knows until we get our hands on them

 

[FrgMstr]

According to the article the limit of 4GB may be due to the physical size of the core and the limit of HBM1 to stacks that are 4 deep.

This is currently the issue. As the production process matures, AMD thinks that it may well be able to stack more DRAM into the Gen I tech. Bottom line is that it has to start somewhere in the real world, and that is what AMD is doing.

 

[Nauseous]

4GB just seems lack luster, i think most of us were expecting 8 at launch. But maybe it will be better than what we all think.

 

[Lunas]

Humm this looks like video cards are going to shrink in size greatly I am curious how the 500mhz clock speed on the hbm will affect the oc on the cores...

Also the 4gb limit is gen 1 hbm didn't nvidia say they were using gen 2 hbm in 2016... and if gen 2 is only that far away that means in 2016 we will be seeing refreshes with more gen2 hbm of amd cards.

Gen 2 said stacks up to 8 high and 8 packages on die... so 16gb? on gen 2 hbm after the manufacture process matures.

so when do we see stacked gpus

 

[Modred189]

Catch 22:
Current GPUs are not bottlenecked by memory bandwidth (much), and are instead themselves the weak link. HOWEVER, with a GPU with enough horsepower to use the greater bandwidth, you'll need more than 4gb of vram.

That said, i could see this being very useful in the next generation of SLI/CF where the memory is not mirrored across the two cards. Having 2x4gb cards totally 8gb of HBM at 4k might actually help in some instances.

 

[Armenius]

Are there any real cons to stacked memory? The only thing i can think of is more power\heat being focused in a smaller area, with standard gddr it was more spread out over a larger area. Still seems odd that the rumours are pointing to these new cards being substantially smaller than what we've had previously. I think most people by now had got used to having to apply for planning permission to install a new card

Cards are still going to be two slots tall, which is my biggest issue. I'd like to see video cards get back to being only one slot.

 

[Modred189]

Cards are still going to be two slots tall, which is my biggest issue. I'd like to see video cards get back to being only one slot.

LIfe will get real interesting for waterblock manufacturers.

 

[cinnamonandgravy]

fuck yeah HBM! OCing VRAM is paltry affair - a solid OC adds what? 50 GB/sec, if youre lucky. HBM is nigged out and ready to sprout.

cant wait for this stuff to become ubiquitous around the 1TB/sec rate - itll be fun to see how engines change.

single card 4k @ ~8k + smaa + msaa were we coooooommmmmeeee.

 

[Weenis]

I thought nvidia was already prototyping this as well?

http://devblogs.nvidia.com/parallelforall/nvlink-pascal-stacked-memory-feeding-appetite-big-data/

 

[Armenius]

I thought nvidia was already prototyping this as well?

http://devblogs.nvidia.com/parallelforall/nvlink-pascal-stacked-memory-feeding-appetite-big-data/

Yes, but AMD cards are the first to production with Fiji launching next month.

 

[Teenyman45]

Hell, I'll be happy for the reduction in package size alone. There are some great m-itx cases around and modern GPUs are a bit too long for the majority of them. If they can make a card in the 200mm range it opens up case options. There are already niche m-itx GPUs, but they are more expensive compared to their std length length siblings.

Oh yes, that size reduction will be a lifesaver in fitting new video cards into my CaseLabs TH-10. Getting 7970's in there was such a tight squeeze.

But, yeah definitely see some utility for m-itx and 1U racks.

 

[TroubleMagnet]

The disadvantage of stacked memory (which has been kicked around since 1996, likely earlier) is it drops your effective yield down and thus drives your costs up. Consider a 5% fallout rate for the parts. Once you stack them you have about a 19% fallout rate since if any one of the parts is bad the whole stack is bad. So if your chance of a good part is R and your stack is S high, your final stacked yield is R^S. So you better have a damm good process if you don't want to throw a bunch of stuff away.

On the power savings, remember you save at least the same amount of power in the GPU as you save in the DRAM since the GPU has to drive the bus too.

If they can get the costs low enough this could be a bigger boost for the mid and even low end cards as they usually get cut down VRAM busses and end up starved for VRAM bandwidth. This could mean even with a cut down memory bus (2 stacks of 4) they won't be bandwidth bottle-necked. Depends on how cheap they can get the interposer to be I bet.

 

[Creig]

Yeah, basically the biggest "con" would seem to be heat. It's all packed into a much smaller, tighter area.

AMD may have already accounted for that if the leaked images of the 390X are close to reality. With such a small card housing a powerful GPU, it would follow that AMD will be producing a waterblock large enough to cool not only the GPU, but the VRMs and the memory as well. That's my best guess, anyhow.

 

[Zarathustra[H]]

The disadvantage of stacked memory (which has been kicked around since 1996, likely earlier) is it drops your effective yield down and thus drives your costs up. Consider a 5% fallout rate for the parts. Once you stack them you have about a 19% fallout rate since if any one of the parts is bad the whole stack is bad. So if your chance of a good part is R and your stack is S high, your final stacked yield is R^S. So you better have a damm good process if you don't want to throw a bunch of stuff away.

Unless you have a way of disabling defective layers in the stack and binning them.

 

[polonyc2]

should be an interesting battle between the 390X with HBM versus the 980 Ti with 6GB VRAM

 

[TroubleMagnet]

Unless you have a way of disabling defective layers in the stack and binning them.

Even if you can salvage the other parts in the stack you're still stuck with a part that has 3/4 (or less) of the capacity that it should. You're likely go to be selling it for half price, if at all. Right now I bet there is zero market for reduced capacity HBM stacks.

 

[Lunas]

Even if you can salvage the other parts in the stack you're still stuck with a part that has 3/4 (or less) of the capacity that it should. You're likely go to be selling it for half price, if at all. Right now I bet there is zero market for reduced capacity HBM stacks.

question is when is the stack assembled before or after the testing. if the stacks of 4 are assembled with 4 known good modules...

 

[LordEC911]

Even if you can salvage the other parts in the stack you're still stuck with a part that has 3/4 (or less) of the capacity that it should. You're likely go to be selling it for half price, if at all. Right now I bet there is zero market for reduced capacity HBM stacks.

Well HBM has a decent amount of built-in redundancy.
Unless you are manufacturing your own HBM, Hynix is selling known good stacks...

Amkor is handling the assembly/packaging, so no risk to AMD.

 

[MisterHipster]

I love AMD, and this is definitely welcome, but we'll see how it all plays out once it hits consumer hands. I have a lot of hope, but I ain't going all-in just yet. Nvidia may have the better approach with wait and see, but I usually like pioneers the best... In my mind, that is the better approach. Anyway, I'm liking where this is going, but short-term we are not going to see a revolution. I'm hoping this entices more people to switch to AMD though.

 

[Mako360]

All I know is that on the Titan X you can overclock, or underclock, the memory by hundreds of MHz (if not a 1000MHz+) and it has near-zero effect on actual fps at 4k. Like less than 1-2fps net gain/loss.

So clearly memory bandwidth isn't a bottleneck at existing levels.

I'm not sure that AMD/Nvidia going apeshit on their marketing trying to convince us that 390X/Pascal with HBM will benefit us is worth anything other than a smaller PCB package size (which is nice of course).

It's all about the GPU architecture it seems, still.