SK Hynix Updates Memory Catalog to Feature GDDR6 and HBM2

[btarunr]

South Korean DRAM and NAND flash giant SK Hynix updated its product catalog to feature its latest GDDR6 memory, besides HBM2. The company had April announced its first GDDR6 memory products. The first GDDR6 memory chips by SK Hynix come in 8 Gb (1 gigabyte) densities, and data-rates of 14 Gbps and 12 Gbps, with DRAM voltages of 1.35V. The company is giving away small quantities of these chips for product development, mass production will commence soon, and bulk availability is slated for Q4-2017. This would mean actual products implementing these chips could be available only by very-late Q4 2017, or Q1-2018.

A graphics card with 14 Gbps GDDR6 memory across a 256-bit memory bus (8 chips) features 448 GB/s of memory bandwidth. A card with 384-bit (12 chips), should have 672 GB/s at its disposal. Likewise, the 12 Gbps memory chips offer 384 GB/s in 256-bit (8-chip) setups, and 576 GB/s in 384-bit (12-chip) setups. Meanwhile, SK Hynix also updated its HBM2 catalog to feature a 32 Gb (4 gigabyte) HBM2 stack, with a clock speed of 1.60 Gbps. The 2.00 Gbps stack which featured in the Q4-2016 version of this catalog is no longer available. At 1.60 Gbps, a GPU with four stacks has 819.2 GB/s of memory bandwidth. A chip with two stacks, such as the purported "Vega 10" prototype that has made several media appearances, hence has 409.6 GB/s.



View at TechPowerUp Main Site

 

[PowerPC]

Man... almost one TB/s. That's insane!

 

[Assimilator]

Man... almost one TB/s. That's insane!

But is it necessary, and does it justify the price? Considering Fury X was primarily held back by memory capacity, not bandwidth, and that even the fastest NVIDIA card (GTX 1080 Ti) can "only" hit 484GB/s, I feel like there isn't really a need for HBM in consumer cards. SK Hynix certainly wouldn't have invested in GDDR6 if they thought HBM was going to succeed it.

 

[Prima.Vera]

But is it necessary, and does it justify the price? Considering Fury X was primarily held back by memory capacity, not bandwidth, and that even the fastest NVIDIA card (GTX 1080 Ti) can "only" hit 484GB/s, I feel like there isn't really a need for HBM in consumer cards. SK Hynix certainly wouldn't have invested in GDDR6 if they thought HBM was going to succeed it.

If the GPU is able to process that much data, sure, even 1TB/s is not enough. Especially on 4K surround setups or 8K future displays.
The question is, are there any CPU, except Xeons, that are able to process that much data coming from the GPU??

 

[Nuckles56]

If the GPU is able to process that much data, sure, even 1TB/s is not enough. Especially on 4K surround setups or 8K future displays.
The question is, are there any CPU, except Xeons, that are able to process that much data coming from the GPU??

The other big issue is getting that data across the PCI-E bus too

 

[springs113]

It's not a fact of is it necessary but...I say bring it on, advancement is needed in all sectors. Isn't the gen4(pci-e) standard due out soon enough(2yrs).
We need the advancement of every field.

 

[PowerPC]

It's not a fact of is it necessary but...I say bring it on, advancement is needed in all sectors. Isn't the gen4(pci-e) standard due out soon enough(2yrs).
We need the advancement of every field.

The exciting part is the direction that they are 'cutting through' with this. Just the fact that they're improving these numbers, even if it may not be something you can take much advantage of in the first gen or so. They're saying "look what we can do" and letting the others hear that they have to catch up or be known for bottlenecking the whole industry. Eventually the bottlenecks go away and these things do become useful. I would even argue that having advancement in this area without "much use" right now only strengthens the argument that this is something we need to be looking at more intensely.