Comments Locked

12 Comments

Back to Article

  • jjj - Thursday, December 21, 2017 - link

    Likely the "increased speed margins" means 8Gb/s for the 8GB GDDR5 die vs 7Gb/s max for the older parts.
  • jjj - Thursday, December 21, 2017 - link

    Nm seems they had 8Gb/s already.
  • menting - Thursday, December 21, 2017 - link

    it most likely means that even though they are speed binned for the max speed bin, they can safely run higher than that speed.
  • Yojimbo - Thursday, December 21, 2017 - link

    During Micron's earnings release conference call on December 19th, Micron's CEO Sanjay Mehrotra said in a prepared remark, "We sampled industry-leading 16Gb-per-second GDDR6 products to key customers and are seeing significant interest in automotive and networking applications that need the high bandwidth this memory provides. We plan to ramp GDDR6 to production in early calendar 2018 for the graphics market, followed by other high-performance applications such as automotive and networking."

    In other words, according to his statement, they are sampling 16 Gb/s GDDR6 even though it's not listed on that web site. Also, the choice of words "We plan to ramp GDDR6 to production in early 2018..." seems a bit stronger to me than "first half 2018". I wouldn't consider May or June early 2018. (Of course, as Tesla shows, different companies have different meanings of the word "production".) So, it looks reasonable that graphics cards based on the next generation architecture could be available by the middle of 2018, even H1 2018.
  • Pork@III - Friday, December 22, 2017 - link

    Where is speed upgrade? Or just new model number?
  • Yojimbo - Friday, December 22, 2017 - link

    Bandwidth must be balanced by power efficiency. If you are willing to pay 150 W for the memory subsystem you can have massive bandwidth. GDDR6 can reach high bandwidths at lower power dissipation than GDDR5. Micron offers GDDR5 up to 8 Gb/s and GDDR5X up to 12 Gb/s. GDDR6 with 16 Gb/s signaling is incoming, apparently with reasonable power usage. So the memory bus doesn't have to be so wide to provide high bandwidths. In fact, it looks like without GDDR6 NVIDIA would have to turn to HBM2 to get enough memory bandwidth at a reasonable power draw for their next generation high-end graphics cards.
  • Pork@III - Friday, December 22, 2017 - link

    "GDDR6 with 16 Gb/s signaling is incoming"
    Mmm but use twice cnannels per half physically I/O
    You must explain to me how this can actually happen in reality, with an increase in bandwidth?
  • 0ldman79 - Saturday, December 23, 2017 - link

    The other side of that is adding in more chips, more channels at the same power envelope.

    Eventually GDDR5 is going to be limited by power, GDDR6 will be able to go farther. Evolution, not revolution...
  • deil - Friday, December 22, 2017 - link

    so when I will get ddr5 for CPU ? :) gib
  • iwod - Friday, December 22, 2017 - link

    As far as we know HBM3 will be based on 7nm, which means it is likely to be late 2018 before we see it.

    GDDR6, assuming cost is still inline with GDDR5, sounds like offer a much better price to performance then HBM3.
  • menting - Monday, December 25, 2017 - link

    7nm is impossible, unless you are talking about the controller only. Samsung is on the 1y node for DRAM, and Micron and Hynix will soon follow with their own 1y node, which is ~16nm approx.
  • Yojimbo - Monday, December 25, 2017 - link

    I think he's referring to Rambus's recent reveal of their plans for HBM3 and DDR5 PHY chips on 7nm. Apparently the PHY sits between the memory controller and the DRAM cells. In HBM2 I am guessing it's in the logic die at the base of each stack, but I don't really know. I'm also guessing it's possible HBM3 could be made with a PHY produced on some other node than 7nm. I just did a quick search for who else might make HBM PHYs. I checked Inphi but apparently Rambus bought Inphi's memory interconnect business in 2016. I have no idea if Inphi made them before that or not. Rambus, which is mainly an IP company, started designing a few low revenue chips some years back because, from what I remember reading the CEO say a year or two ago, their customers weren't happy with what was available on the market at the time.

    In any case, 7nm or not, I doubt we'll see HBM3 before 2019. It seems that in August 2016 Samsung promised 2019 as the likely year of introduction of their HBM3 chips. That's more likely to get pushed back than moved forward.

Log in

Don't have an account? Sign up now