Howdy, Stranger!

It looks like you're new here. If you want to get involved, click one of these buttons!

Hynix announces that they're building GDDR6 memory for a high-end video card in early 2018.

QuizzicalQuizzical Member EpicPosts: 18,399
edited April 2017 in Hardware

And probably enough specs on it to make someone mad.  Apparently the unspecified upcoming GPU has a 384-bit memory bus, with the memory rated at 16 Gbps, for 768 GB/s of memory bandwidth.  Hynix says that it's coming in "early 2018".

Nvidia has made lots of GPUs with a 384-bit memory bus, including the GeForce GTX 480, GTX 580, GTX 780 Ti, GTX 980 Ti, and GTX 1080 Ti.  So one could have guessed a top end GeForce card with a 384-bit GDDR6 memory bus would show up at some point.  That vague guess wouldn't tell you the timing of the launch, however.

AMD, meanwhile, nearly always goes with powers of 2 for their memory bus width.  They've only made one GPU ever with a 384-bit memory bus: the Tahiti chip of the Radeon HD 7970/7950.  And with AMD having already announced a move to HBM2 for their high end, to follow that with GDDR6 would seem like a downgrade.

Edit:  here's the press release straight form the original source:
Post edited by Quizzical on


  • HrimnirHrimnir Member RarePosts: 2,392
    My guess is it's gonna be Nvidia, I just don't see why AMD would move away from HBM/2 on their high end parts.

    "The surest way to corrupt a youth is to instruct him to hold in higher esteem those who think alike than those who think differently."

    - Friedrich Nietzsche

  • bhugbhug Member UncommonPosts: 931
    gddr6 2q18 @ 16Gbps, sk hynix, (768 GB/s in 384b I/O bus); nvidia volta

    HighBandwidthMemory (hbm stacks memory chips on top of each other around a central core, silicon vias [TSV] wires connect stacks, 2.5D architecture, emphasizes low clock rates and extremely wide bus) gddr5 refresh 201512; hbm2 1.2v IO/core v, 2-4-8 stack height, 256Gbps, overpriced;

     samsung theoretical hbm3 (high bandwidth + low cost) 2020 (? industry stability) vs gddr6; needs material that can tolerate +temp in large-scale die stacking; 10nm and 7nm nodes in EUV lithography

     ? vs micron ddr5 gddr5x (2D planar Silicon, 2q16) plans;


  • TorvalTorval Member LegendaryPosts: 15,751

    Hrimnir said:

    My guess is it's gonna be Nvidia, I just don't see why AMD would move away from HBM/2 on their high end parts.

    Not for the high end anyway. I could see mid-grade and lower cards staying with GDDR because it should get cheaper and it's very familiar.

    How closely can GDDR-6 compete with HBM-2?

    Here is part of a post I made this January when the 5X details were published. GDDR is at end of life without some major change and then it would be GDDR in pretty much name only. I could see going with GDDR6 for cost maybe in a respin?

    From Jan 2016
    JEDEC has published final specs for both GDDR5X and HBM2. Here are the links for detailed information:



    HBM2 outperforms GDDR5X by a good margin, but is also likely a lot more expensive. GDDR5X performs nearly as good as HBM (about 12% below it). HBM2 is already in production while GDDR5X is about 6 months out.

    One advantage 5X has over HBM2 is that it closely mirrors the external implementation of GDDR5 and takes very little modification to implement. One huge advantage HBM2 has over 5X is that it's already in production. The article about 5X points out some pretty relevant points about memory consumption and performance. The HBM2 article shows just how much more powerful HBM2 is and how the technology can grow while GDDR is in a very mature stage of life where it doesn't have a lot of growth room.

Sign In or Register to comment.