SK Hynix & Samsung Discuss 'Low-Cost HBM', HBM2 & HBM3

[HKZeroFive]

 

Quote

The low cost HBM solution is presented by Samsung and is explained to be more cost effective. It is faster than HBM1 but slower than HBM2 however, it’s supposed to be significantly cheaper. Comparing both HBM2 and HBM (low cost), it is seen that the latter comes with lesser TSVs. TSV standards for through silicon vias which are used to handle the I/O on the DRAM die. The number is reduced from 1024 on an HBM2 stack to 512 on a low cost stack.

 

The end result is a faster pin speed of 3 GB/s (+) that can deliver 200 GB/s compared to 256 GB/s on HBM2. The lower 512-bit interface across 2 / 4 stacks would equate to 1024 / 2048. Samsung believes that they can easily produce these chips in larger quantity and ship it to a mass market.

 

Quote

With HBM2 hitting production, SK Hynix and Samsung are already prepping up for the next iteration of HBM memory as demand for bandwidth and efficiency is ever increasing. SK Hynix terms their next solution as HBM3 or HBMx while Samsung calls it xHBM or Extreme HBM. The specifications for HBM3 have not been finalized yet and are mostly under consideration at this moment. HBM3 would offer twice the bandwidth and feature a very attractive price. We are talking about 512 GB/s bandwidth from these chips compared to 256 GB/s offered on HBM2. Four of these stacks would result in over 2 TB/s of bandwidth. Some points in consideration for the next-gen HBM are cost-effectiveness, form factors, power, density and bandwidth. Currently, HBM2 can go as high as 48 GB in terms of capacity so probably expect around 64 GB when HBM3 arrives.

 

Some real interesting stuff to see HBM3 already in development when HBM2 is still a relatively new technology that was only just recently introduced to the server market, whilst there are no offerings of such consumer products. However, I'd see cost to be a huge obstacle. HBM and HBM2 proved to be expensive so this shouldn't be a huge surprise to anyone. Speaking of cost, the mention of low-cost (I would guess that's subjective) HBM being faster HBM1 does sound exciting and could potentially be released to wide-spread market. But let's see if anyone pushes it.

 

This talk of DDR5 and HBM3 indicates that technology is indeed moving at a fast rate. In my eyes, that's a good thing.

 

Sources:

 

https://www.computerbase.de/2016-08/sk-hynix-samsung-hbm2-hbm3/

http://wccftech.com/sk-hynix-samsung-micron-hbm-hmc-ddr5-hot-chips/

[Curufinwe_wins]

Makes sense to try to make a lower cost version, but the same issue of yields is going to hurt them a lot.

[zMeul]

it still needs to come in packages that allow more than 4GB 

[PRTI]

5 minutes ago, zMeul said:

it still needs to come in packages that allow more than 4GB 

They will probably add 3gb just to mock out 4k  

[zMeul]

Just now, keNNySOC said:

4K UHD

¯\_(ツ)_/¯

[BuckGup]

11 minutes ago, zMeul said:

it still needs to come in packages that allow more than 4GB 

HBM2 allows for 4gb stacks so you can get 16gb

[zMeul]

2 minutes ago, SamStrecker said:

HBM2 allows for 4gb stacks so you can get 16gb

I was comparing to HBM1 that only allows for 1GB stacks; 4 stacks -> 4GB - that's it

 

and HBM2 is seriously expensive, I doubt we'll see it on consumer grade products anytime soon

[BuckGup]

1 minute ago, zMeul said:

I was comparing to HBM1 that only allows for 1GB stacks; 4 stacks -> 4GB - that's it

Yeah but who would use HBM1 when we have HBM2 available 

[DocSwag]

16 minutes ago, rrubberr said:

Who would use either in the first place?

Hbm uses less power, takes up less space, and tends to be able to offer more bandwidth. The only drawback is price.

 

There's a reason nvidia's using hbm2 in the p100 and AMD will use hbm2 in vega.

[DocSwag]

2 minutes ago, rrubberr said:

Yeah, and as the latest cut-down Titan tells us, cost is everything. What's the point of manufacturing computer parts if you can't scalp them for 20x what they cost to make? HBM in P100 is supposedly because HPC applications need to be able to access memory that quickly. Other than that, I'm not sure that BF4 really needs to move 300 GB of anything through VRAM in a second.

Right now for GPUs it isn't as much about bandwidth as it is power consumption. HBM2 costs more than something like gddr5x but it also draws less power as well as taking up less space. It doesn't cost vastly more... It's probably at most two or three times more. And the reduced power consumption means increased performance-if Fiji didn't have hbm fury X would either have had to draw 325 watts or something or had to settle for fury performance and fury would have been under GTX 980. There will be some need for the extra bandwidth-the Titan X already has 480 gbps and to go any further you'd have to go hbm2 or a 512 bit bus but there definitely is no gaming GPU right now that needs 1 TB/s. HBM in general is just a trade off. If you want to reach even higher levels of performance you'll have to pay a little more for it. That's how nearly everything works and hbm is no exception.

[Curufinwe_wins]

17 hours ago, DocSwag said:

Right now for GPUs it isn't as much about bandwidth as it is power consumption. HBM2 costs more than something like gddr5x but it also draws less power as well as taking up less space. It doesn't cost vastly more... It's probably at most two or three times more. And the reduced power consumption means increased performance-if Fiji didn't have hbm fury X would either have had to draw 325 watts or something or had to settle for fury performance and fury would have been under GTX 980. There will be some need for the extra bandwidth-the Titan X already has 480 gbps and to go any further you'd have to go hbm2 or a 512 bit bus but there definitely is no gaming GPU right now that needs 1 TB/s. HBM in general is just a trade off. If you want to reach even higher levels of performance you'll have to pay a little more for it. That's how nearly everything works and hbm is no exception.

That isn't the only reason.... HBM has much much worse latency than GDDR5x (like 35ns compared to 70+ns), there could very well be applications where it is actually worse than GDDR5x on a sufficiently wide bus. Even if you eat the power consumption hit of GDDR5x (which BTW HBM's power consumption goes through the roof with increased interface speed, easily expected to exceed GDDR5 at 2GHz), a lower latency, cheaper solution that also doesn't require extreme TSV binning (a single HBM failure means throwing out the die, whereas GDDR5(x) is BGA) is kinda a win-win for GPU companies. Assuming they can give sufficient bandwidth, which currently seems to be the case.

[patrickjp93]

18 hours ago, DocSwag said:

Hbm uses less power, takes up less space, and tends to be able to offer more bandwidth. The only drawback is price.

 

There's a reason nvidia's using hbm2 in the p100 and AMD will use hbm2 in vega.

It uses just as much power. Stop being so easily fooled. It's also more expensive because of the manufacturing restraints put on both the GPU and the HBM itself. Interposers are not cheap.

[Coaxialgamer]

1 hour ago, patrickjp93 said:

It uses just as much power. Stop being so easily fooled. It's also more expensive because of the manufacturing restraints put on both the GPU and the HBM itself. Interposers are not cheap.

true , but hbm is vastly more efficent . While you might have both a gddr5x and hbm2 setup drawing ~20w , the hbm2 will provide 2+ times the bandwidth 

[Prysin]

20 hours ago, zMeul said:

it still needs to come in packages that allow more than 4GB 

not if it is cheap enough to be used with APUs or consoles.

[patrickjp93]

31 minutes ago, Coaxialgamer said:

true , but hbm is vastly more efficent . While you might have both a gddr5x and hbm2 setup drawing ~20w , the hbm2 will provide 2+ times the bandwidth 

No. 512-bit bus at 16GHz effective is about 25W and is a full 1TB/s. That's slightly over what 32GB of HBM2 at 1TB/s would use, but HBM2's problem is it becomes exceedingly power hungry as you push the clocks. Wccf had an article where iirc HBM2 is shown to have 35W power draw at just 1.2TB/s.

[Curufinwe_wins]

3 hours ago, patrickjp93 said:

It uses just as much power. Stop being so easily fooled. It's also more expensive because of the manufacturing restraints put on both the GPU and the HBM itself. Interposers are not cheap.

As you well know it isn't so much the interposer is very expensive (although it isnt cheap), but the fact that you don't know if the chip works till after it has been attached to the interposer and die means that in many cases on failed HBM chip requires crapping the entire gpu die as well (and all other hbm chips), which on a per unit basis adds a huge amount of cost.

[DocSwag]

1 hour ago, patrickjp93 said:

No. 512-bit bus at 16GHz effective is about 25W and is a full 1TB/s. That's slightly over what 32GB of HBM2 at 1TB/s would use, but HBM2's problem is it becomes exceedingly power hungry as you push the clocks. Wccf had an article where iirc HBM2 is shown to have 35W power draw at just 1.2TB/s.

We still don't have 16 gbps gddr5x. In fact the best we're getting right now is 10gbps.

 

Also, wccftech? Ew....

[patrickjp93]

2 hours ago, Curufinwe_wins said:

As you well know it isn't so much the interposer is very expensive (although it isnt cheap), but the fact that you don't know if the chip works till after it has been attached to the interposer and die means that in many cases on failed HBM chip requires crapping the entire gpu die as well (and all other hbm chips), which on a per unit basis adds a huge amount of cost.

You can test HBM stacks before attaching to the final interposer. That doesn't mean the attachment process can't introduce a failure, but it's not quite THAT bad.

 

2 hours ago, DocSwag said:

We still don't have 16 gbps gddr5x. In fact the best we're getting right now is 10gbps.

 

Also, wccftech? Ew....

Wccf wasn't the original source. The source was a paper by an electrical engineer. Wccf is just the only one who made an article. And we have 14Gbps GDDR5X already.

[BuckGup]

On August 21, 2016 at 4:48 PM, rrubberr said:

Who would use either in the first place?

Ludicrous bandwidth for high res textures like 4K

[patrickjp93]

8 minutes ago, SamStrecker said:

Ludicrous bandwidth for high res textures like 4K

Bandwidth is not th limiting factor in that case. Latency is. The time to load one object/texture for manipulation from memory is a latency problem. The time to load many of them from memory back to back is a bandwidth problem. Assuming good optimization, you can hide the latency of the 3rd texture load and on, but that doesn't change the fact latency is the limiting factor even for GPUs now.

[TheRandomness]

On 22 August 2016 at 8:06 AM, DocSwag said:

Hbm uses less power, takes up less space, and tends to be able to offer more bandwidth. The only drawback is price.

 

There's a reason nvidia's using hbm2 in the p100 and AMD might use hbm2 in vega.

ftfy

[BuckGup]

1 hour ago, patrickjp93 said:

Bandwidth is not th limiting factor in that case. Latency is. The time to load one object/texture for manipulation from memory is a latency problem. The time to load many of them from memory back to back is a bandwidth problem. Assuming good optimization, you can hide the latency of the 3rd texture load and on, but that doesn't change the fact latency is the limiting factor even for GPUs now.

Would HBM reduce latency by the sheer fact of trace length? 

[patrickjp93]

15 minutes ago, SamStrecker said:

Would HBM reduce latency by the sheer fact of trace length? 

Not much vs. GDDR5. It's still around 40ns vs. 20 ish for DDR4, and 7ns for HMC. Someone posted a table around here yesterday. Curufinwe and @Prysin probably know where it is.

[Doobeedoo]

Well great to see this tech going forward and also scales lovely with in all improvement corners all together. While HBM3 is far off no doubt, can't wait to see next flagship GPU with HBM2 now. Vega I guess.

[Lethal Seraph]

Time to find a buyer for my kidney, I guess.