SK Hynix confirms GDDR6 in 2018

[Hunter259]

Why do people care so much about what kind of vram is in their GPU's? Vram speed rarely is ever a bottleneck as has been shown time and time again.

[Dylanc1500]

16 minutes ago, NumLock21 said:

Yeah but the customers on that side don't know hbm is made by amd.

 

Partly they developed in partnership with Hynix the first gen HBM. Second Gen was collaborated with multiple companies with Samsung and Hynix being two of the manufacturers.

[Godlygamer23]

1 hour ago, Hunter259 said:

Why do people care so much about what kind of vram is in their GPU's? Vram speed rarely is ever a bottleneck as has been shown time and time again.

There's another factor to consider with HBM - space taken up. Not only is the bandwidth higher, but it's also integrated in with the GPU package, allowing for potentially shorter PCBs, or more robust, larger components to be placed on it. 

[Hunter259]

53 minutes ago, Godlygamer23 said:

There's another factor to consider with HBM - space taken up. Not only is the bandwidth higher, but it's also integrated in with the GPU package, allowing for potentially shorter PCBs, or more robust, larger components to be placed on it. 

It also causes increased complexity in the GPU Die which could increase cost and decrease reliability in the short term. For desktop GPU's, I don't see much point. Mobile is a completely different story however.

[Godlygamer23]

2 hours ago, Hunter259 said:

It also causes increased complexity in the GPU Die which could increase cost and decrease reliability in the short term. For desktop GPU's, I don't see much point. Mobile is a completely different story however.

Which is a valid argument, but it's not inherently just about bandwidth. 

[Dabombinable]

1 hour ago, Godlygamer23 said:

Which is a valid argument, but it's not inherently just about bandwidth. 

Power consumption is another big plus with HBM.

[VagabondWraith]

5 hours ago, Hunter259 said:

Why do people care so much about what kind of vram is in their GPU's? Vram speed rarely is ever a bottleneck as has been shown time and time again.

Cos people are dumb. Even if G6 memory is more than sufficient they'll get mad cos no HBM. I don't care if I have SDRAM as long as it's fast enough to keep the GPU fed. The only real advantage to HBM is smaller surface area used for smaller GPU's.

[SpaceGhostC2C]

13 hours ago, AluminiumTech said:

HBM2 probably costs similar amounts to GDDR5X.

Mmm. I have systematically heard/read that HBM was more expensive, hence why GDDR5/GDDR5X was used everywhere except at the highest end. Are you sure of that?

[Dabombinable]

1 minute ago, SpaceGhostC2C said:

Mmm. I have systematically heard/read that HBM was more expensive, hence why GDDR5/GDDR5X was used everywhere except at the highest end. Are you sure of that?

Considering the way it has to be implemented plus the low availability plus the cost of the few graphics cards that use it-it definitely cost more than GDDR5 and GDDR5X. Part of the low availability however would be due to use outside of gaming in compute orientated cards and other products which reuire high memory bandwidth with low power consumption. 

[Mira Yurizaki]

I would imagine including HBM decreases the yield rate. Or at the very least, it doesn't allow for much flexibility. Who knows, HBM could be the future, but at the moment, I'm thinking it's only useful on the top end and will only trickle its way down once GPU designers hit a wall trying to squeeze everything they can out of GDDR.

 

A lot of people were all "lol, nvidia, ur drunk, go home" when they saw the GTX 980 having a "measely" 256-bit bus while AMD was touting their 4096-bit HBM bus on the R9 Fury. I guess it's like they say, it's not the size of the thing, but how you use it.

[Dabombinable]

1 minute ago, M.Yurizaki said:

I would imagine including HBM decreases the yield rate. Or at the very least, it doesn't allow for much flexibility. Who knows, HBM could be the future, but at the moment, I'm thinking it's only useful on the top end and will only trickle its way down once GPU designers hit a wall trying to squeeze everything they can out of GDDR.

 

A lot of people were all "lol, nvidia, ur drunk, go home" when they saw the GTX 980 having a "measely" 256-bit bus while AMD was touting their 4096-bit HBM bus on the R9 Fury. I guess it's like they say, it's not the size of the thing, but how you use it.

And with Nvidia having better compression...

[AlTech]

48 minutes ago, M.Yurizaki said:

I would imagine including HBM decreases the yield rate. Or at the very least, it doesn't allow for much flexibility. Who knows, HBM could be the future, but at the moment, I'm thinking it's only useful on the top end and will only trickle its way down once GPU designers hit a wall trying to squeeze everything they can out of GDDR.

HBM2 can scale better in terms of being more available for mid range products.

 

Since the new offerings with 8GB HBM2 only requires 2 DRAM modules out of 4.  And for 4GB cards this only means a single DRAM module. Significantly cheaper than a 4096 bit fully enabled package.